作家、未来学家

AUTHOR, FUTURIST

人工智能正在迅速从科幻小说领域转变为我们日常生活的现实。我们的设备能够理解我们所说的话、与我们交谈，并以越来越流畅的方式在多种语言之间进行翻译。人工智能驱动的视觉识别算法的表现优于人类，并开始在从自动驾驶汽车到诊断医学图像中的癌症的系统等各个领域得到应用。主要媒体组织越来越依赖自动化新闻报道，将原始数据转化为连贯的新闻报道，这些报道与人类记者撰写的报道几乎没有区别。

Artificial intelligence is rapidly transitioning from the realm of science fiction to the reality of our daily lives. Our devices understand what we say, speak to us, and translate between languages with ever-increasing fluency. AI-powered visual recognition algorithms are outperforming people and beginning to find applications in everything from self-driving cars to systems that diagnose cancer in medical images. Major media organizations increasingly rely on automated journalism to turn raw data into coherent news stories that are virtually indistinguishable from those written by human journalists.

这个清单可以列很多，而且越来越明显的是，人工智能有望成为塑造我们世界的最重要力量之一。与更专业的创新不同，人工智能正在成为一种真正的通用技术。换句话说，它正在演变成一种公用事业——与电力类似——最终可能会扩展到每个行业、我们经济的每个部门以及科学、社会和文化的几乎每个方面。

The list goes on and on, and it is becoming evident that AI is poised to become one of the most important forces shaping our world. Unlike more specialized innovations, artificial intelligence is becoming a true general-purpose technology. In other words, it is evolving into a utility—not unlike electricity—that is likely to ultimately scale across every industry, every sector of our economy, and nearly every aspect of science, society and culture.

过去几年，人工智能所展现出的强大力量引起了媒体的广泛关注和评论。无数的新闻文章、书籍、纪录片和电视节目不遗余力地列举人工智能的成就，并预示着一个新时代的到来。结果，这种说法有时令人费解，既有谨慎的、基于证据的分析，也有炒作、猜测和完全可以称之为恐吓的言论。我们被告知，完全自动驾驶汽车将在短短几年内与我们共享道路，而卡车、出租车和优步司机的数百万个工作岗位即将消失。某些机器学习算法中已经发现了种族和性别偏见的证据，而对面部识别等人工智能技术将如何影响隐私的担忧似乎是有道理的。媒体经常报道机器人将很快被武器化，或者真正智能（或超级智能）的机器有朝一日可能对人类构成生存威胁的警告。许多非常著名的公众人物（其中没有一个是真正的人工智能专家）都发表了看法。伊隆·马斯克的言论尤其极端，他宣称人工智能研究正在“召唤恶魔”，“人工智能比核武器更危险”。甚至包括亨利·基辛格和已故的斯蒂芬·霍金在内的一些较为温和的人士也发出了可怕的警告。

The demonstrated power of artificial intelligence has, in the last few years, led to massive media exposure and commentary. Countless news articles, books, documentary films and television programs breathlessly enumerate AI’s accomplishments and herald the dawn of a new era. The result has been a sometimes incomprehensible mixture of careful, evidence-based analysis, together with hype, speculation and what might be characterized as outright fear-mongering. We are told that fully autonomous self-driving cars will be sharing our roads in just a few years—and that millions of jobs for truck, taxi and Uber drivers are on the verge of vaporizing. Evidence of racial and gender bias has been detected in certain machine learning algorithms, and concerns about how AI-powered technologies such as facial recognition will impact privacy seem well-founded. Warnings that robots will soon be weaponized, or that truly intelligent (or superintelligent) machines might someday represent an existential threat to humanity, are regularly reported in the media. A number of very prominent public figures—none of whom are actual AI experts—have weighed in. Elon Musk has used especially extreme rhetoric, declaring that AI research is “summoning the demon” and that “AI is more dangerous than nuclear weapons.” Even less volatile individuals, including Henry Kissinger and the late Stephen Hawking, have issued dire warnings.

本书旨在通过与世界上一些最杰出的人工智能研究科学家和企业家进行一系列深入、广泛的对话，阐明人工智能领域及其相关的机遇和风险。其中许多人做出了开创性的贡献，直接影响了我们周围所见的变革；其他人则创立了推动人工智能、机器人和机器学习前沿发展的公司。

The purpose of this book is to illuminate the field of artificial intelligence—as well as the opportunities and risks associated with it—by having a series of deep, wide-ranging conversations with some of the world’s most prominent AI research scientists and entrepreneurs. Many of these people have made seminal contributions that directly underlie the transformations we see all around us; others have founded companies that are pushing the frontiers of AI, robotics and machine learning.

当然，挑选出某一领域最杰出、最有影响力的人物名单是一项主观工作，毫无疑问，还有很多其他人已经或正在为人工智能的发展做出重要贡献。尽管如此，我相信，如果你让几乎任何一位对该领域有深入了解的人列出一份塑造当代人工智能研究的最重要人物名单，你得到的名单将与本书采访的个人大体相同。我在这里列出的这些人确实是机器智能的缔造者——而且，从广义上讲，他们也是机器智能即将引发的革命的缔造者。

Selecting a list of the most prominent and influential people working in a field is, of course, a subjective exercise, and without doubt there are many other people who have made, or are making, critical contributions to the advancement of AI. Nonetheless, I am confident that if you were to ask nearly anyone with a deep knowledge of the field to compose a list of the most important minds who have shaped contemporary research in artificial intelligence, you would receive a list of names that substantially overlaps with the individuals interviewed in this book. The men and women I have included here are truly the architects of machine intelligence—and, by extension, of the revolution it will soon unleash.

这里记录的对话一般都是开放式的，但旨在解决人工智能不断发展过程中我们面临的一些最紧迫的问题：哪些特定的人工智能方法和技术最有前景，未来几年我们可能会看到什么样的突破？真正的思考机器——或者人类级别的人工智能——真的有可能出现吗？这种突破多久会出现？我们应该真正担心与人工智能相关的哪些风险或威胁？我们应该如何解决这些问题？政府监管是否发挥了作用？人工智能会引发大规模的经济和就业市场混乱吗？还是这些担忧被夸大了？超级智能机器有朝一日会摆脱我们的控制，构成真正的威胁吗？我们是否应该担心人工智能“军备竞赛”，或者其他实行威权政治制度的国家，尤其是中国，最终可能会占据主导地位？

The conversations recorded here are generally open-ended, but are designed to address some of the most pressing questions that face us as artificial intelligence continues to advance: What specific AI approaches and technologies are most promising, and what kind of breakthroughs might we see in the coming years? Are true thinking machines—or human-level AI—a real possibility and how soon might such a breakthrough occur? What risks, or threats, associated with artificial intelligence should we be genuinely concerned about? And how should we address those concerns? Is there a role for government regulation? Will AI unleash massive economic and job market disruption, or are these concerns overhyped? Could superintelligent machines someday break free of our control and pose a genuine threat? Should we worry about an AI “arms race,” or that other countries with authoritarian political systems, particularly China, may eventually take the lead?

毋庸置疑，没有人真正知道这些问题的答案。没有人能预测未来。然而，我在这里采访过的人工智能专家确实比其他任何人都更了解当前的技术状态以及即将出现的创新。他们往往拥有数十年的经验，并在创造现在开始展开的革命中发挥了重要作用。因此，他们的想法和意见值得重视。除了我对人工智能领域及其未来的问题之外，我还深入研究了这些人的背景、职业轨迹和当前的研究兴趣，我相信他们不同的出身和不同的成名之路将使阅读变得引人入胜和鼓舞人心。

It goes without saying that no one really knows the answers to these questions. No one can predict the future. However, the AI experts I’ve spoken to here do know more about the current state of the technology, as well as the innovations on the horizon, than virtually anyone else. They often have decades of experience and have been instrumental in creating the revolution that is now beginning to unfold. Therefore, their thoughts and opinions deserve to be given significant weight. In addition to my questions about the field of artificial intelligence and its future, I have also delved into the backgrounds, career trajectories and current research interests of each of these individuals, and I believe their diverse origins and varied paths to prominence will make for fascinating and inspiring reading.

人工智能是一个广泛的研究领域，包含许多分支学科，本文采访的许多研究人员都曾在多个领域工作过。有些人还在其他领域拥有丰富的经验，例如人类认知研究。尽管如此，以下只是简要地尝试创建一个非常粗略的路线图，以展示本文采访的个人与人工智能研究中最重要的最新创新以及未来挑战之间的关系。关于每个人的更多背景信息可以在采访后立即找到的其个人简介中找到。

Artificial intelligence is a broad field of study with a number of subdisciplines, and many of the researchers interviewed here have worked in multiple areas. Some also have deep experience in other fields, such as the study of human cognition. Nonetheless, what follows is a brief attempt to create a very rough road map showing how the individuals interviewed here relate to the most important recent innovations in AI research and to the challenges that lie ahead. More background information about each person is available in his or her biography, which is located immediately after the interview.

过去十年左右，我们看到的绝大多数重大进步——从图像和面部识别，到语言翻译，再到 AlphaGo 征服古老的围棋游戏——都由深度学习或深度神经网络技术推动。人工神经网络至少可以追溯到 20 世纪 50 年代，其中的软件大致模拟了大脑中生物神经元的结构和相互作用。这些网络的简单版本能够执行基本的模式识别任务，并且在早期引起了研究人员的极大热情。然而，到了 20 世纪 60 年代——至少部分是由于人工智能早期先驱之一马文·明斯基对该技术的批评——神经网络失宠并几乎被完全抛弃，因为研究人员转向了其他方法。

The vast majority of the dramatic advances we’ve seen over the past decade or so—everything from image and facial recognition, to language translation, to AlphaGo’s conquest of the ancient game of Go—are powered by a technology known as deep learning, or deep neural networks. Artificial neural networks, in which software roughly emulates the structure and interaction of biological neurons in the brain, date back at least to the 1950s. Simple versions of these networks are able to perform rudimentary pattern recognition tasks, and in the early days generated significant enthusiasm among researchers. By the 1960s, however—at least in part as the direct result of criticism of the technology by Marvin Minsky, one of the early pioneers of AI—neural networks fell out of favor and were almost entirely dismissed as researchers embraced other approaches.

从 20 世纪 80 年代开始的大约 20 年时间里，一小部分研究科学家继续相信并推动神经网络技术的发展。其中最杰出的是 Geoffrey Hinton、Yoshua Bengio 和 Yann LeCun。这三人不仅对深度学习背后的数学理论做出了开创性的贡献，而且还是这项技术的主要传播者。他们共同改进了构建更复杂（或“深度”）网络的方法，这些网络具有多层人工神经元。Hinton、Bengio 和 LeCun 有点像保存和复制古典文本的中世纪僧侣，他们带领神经网络度过了自己的黑暗时代——直到数十年来计算能力的指数级发展，加上可用数据量的几乎不可思议的增长，最终促成了“深度学习复兴”。2012 年，这一进步成为了一场彻底的革命，当时来自多伦多大学的 Hinton 的研究生团队参加了一场大型图像识别比赛，并使用深度学习击败了对手。

Over a roughly 20-year period beginning in the 1980s, a very small group of research scientists continued to believe in and advance the technology of neural networks. Foremost among these were Geoffrey Hinton, Yoshua Bengio and Yann LeCun. These three men not only made seminal contributions to the mathematical theory underlying deep learning, they also served as the technology’s primary evangelists. Together they refined ways to construct much more sophisticated—or “deep”—networks with many layers of artificial neurons. A bit like the medieval monks who preserved and copied classical texts, Hinton, Bengio and LeCun ushered neural networks through their own dark age—until the decades-long exponential advance of computing power, together with a nearly incomprehensible increase in the amount of data available, eventually enabled a “deep learning renaissance.” That progress became an outright revolution in 2012, when a team of Hinton’s graduate students from the University of Toronto entered a major image recognition contest and decimated the competition using deep learning.

在接下来的几年里，深度学习变得无处不在。谷歌、Facebook、微软、亚马逊、苹果以及百度和腾讯等中国领先企业等所有大型科技公司都对这项技术进行了巨额投资，并将其运用到自己的业务中。设计微处理器和图形（或 GPU）芯片的公司，如 NVIDIA 和英特尔，也看到了业务转型，因为他们争相打造针对神经网络优化的硬件。深度学习——至少到目前为止——是推动人工智能革命的主要技术。

In the ensuing years, deep learning has become ubiquitous. Every major technology company—Google, Facebook, Microsoft, Amazon, Apple, as well as leading Chinese firms like Baidu and Tencent—have made huge investments in the technology and leveraged it across their businesses. The companies that design microprocessor and graphics (or GPU) chips, such as NVIDIA and Intel, have also seen their businesses transformed as they rush to build hardware optimized for neural networks. Deep learning—at least so far—is the primary technology that has powered the AI revolution.

本书包括与三位深度学习先驱 Hinton、LeCun 和 Bengio 以及其他几位处于该技术前沿的杰出研究人员的对话。Andrew Ng、Fei-Fei Li、Jeff Dean 和 Demis Hassabis 都在网络搜索、计算机视觉、自动驾驶汽车和更通用的智能等领域开发了先进的神经网络。他们也是围绕深度学习技术开展教学、管理研究机构和创业的公认领导者。

This book includes conversations with the three deep learning pioneers, Hinton, LeCun and Bengio, as well as with several other very prominent researchers at the forefront of the technology. Andrew Ng, Fei-Fei Li, Jeff Dean and Demis Hassabis have all advanced neural networks in areas like web search, computer vision, self-driving cars and more general intelligence. They are also recognized leaders in teaching, managing research organizations, and entrepreneurship centered on deep learning technology.

本书的其余对话一般都是与那些可能被称为深度学习不可知论者，甚至是批评者的人进行的。所有人都承认深度神经网络在过去十年中取得了非凡的成就，但他们可能会认为深度学习只是“工具箱中的一种工具”，要继续取得进展，就需要整合人工智能其他领域的思想。其中一些人，包括 Barbara Grosz 和 David Ferrucci，主要关注理解自然语言的问题。Gary Marcus 和 Josh Tenenbaum 在其职业生涯的大部分时间里致力于研究人类认知。其他人，包括 Oren Etzioni、Stuart Russell 和 Daphne Koller，是人工智能通才，或专注于使用概率技术。在最后一组中，尤迪亚·珀尔 (Judea Pearl) 尤为杰出，他于 2012 年获得了图灵奖（本质上是计算机科学的诺贝尔奖），很大程度上是因为他在人工智能和机器学习中的概率（或贝叶斯）方法方面的工作。

The remaining conversations in this book are generally with people who might be characterized as deep learning agnostics, or perhaps even critics. All would acknowledge the remarkable achievements of deep neural networks over the past decade, but they would likely argue that deep learning is just “one tool in the toolbox” and that continued progress will require integrating ideas from other spheres of artificial intelligence. Some of these, including Barbara Grosz and David Ferrucci, have focused heavily on the problem of understanding natural language. Gary Marcus and Josh Tenenbaum have devoted large portions of their careers to studying human cognition. Others, including Oren Etzioni, Stuart Russell and Daphne Koller, are AI generalists or have focused on using probabilistic techniques. Especially distinguished among this last group is Judea Pearl, who in 2012 won the Turing Award—essentially the Nobel Prize of computer science—in large part for his work on probabilistic (or Bayesian) approaches in AI and machine learning.

除了这种根据对深度学习的态度而划分的非常粗略的领域之外，我采访的几位研究人员还专注于更具体的领域。罗德尼·布鲁克斯 (Rodney Brooks)、丹妮拉·鲁斯 (Daniela Rus) 和辛西娅·布雷齐尔 (Cynthia Breazeal) 都是机器人领域公认的领军人物。布雷齐尔和拉娜·埃尔·卡利乌比 (Rana El Kaliouby) 是构建理解和响应情绪的系统的先驱，因此这些系统具有与人进行社交互动的能力。布莱恩·约翰逊 (Bryan Johnson) 创立了一家初创公司 Kernel，希望最终利用技术来增强人类的认知能力。

Beyond this very rough division defined by their attitude toward deep learning, several of the researchers I spoke to have focused on more specific areas. Rodney Brooks, Daniela Rus and Cynthia Breazeal are all recognized leaders in robotics. Breazeal along with Rana El Kaliouby are pioneers in building systems that understand and respond to emotion, and therefore have the ability to interact socially with people. Bryan Johnson has founded a startup company, Kernel, which hopes to eventually use technology to enhance human cognition.

我认为有三个领域非常受关注，因此每次谈话我都会深入探讨。第一个领域涉及人工智能和机器人对就业市场和经济的潜在影响。我个人的观点是，随着人工智能逐渐证明能够自动化几乎所有常规、可预测的任务（无论是蓝领还是白领），我们将不可避免地看到不平等加剧，甚至可能出现彻底失业，至少在某些工人群体中是如此。我在 2015 年出版的《机器人崛起：技术与失业未来的威胁》一书中阐述了这一观点。

There are three general areas that I judged to be of such high interest that I delved into them in every conversation. The first of these concerns the potential impact of AI and robotics on the job market and the economy. My own view is that as artificial intelligence gradually proves capable of automating nearly any routine, predictable task—regardless of whether it is blue or white collar in nature—we will inevitably see rising inequality and quite possibly outright unemployment, at least among certain groups of workers. I laid out this argument in my 2015 book, Rise of the Robots: Technology and the Threat of a Jobless Future.

我采访过的人对这种潜在的经济混乱以及可能解决这一问题的政策解决方案提出了各种观点。为了更深入地探讨这个话题，我咨询了麦肯锡全球研究所主席詹姆斯·曼尼卡。曼尼卡是一位经验丰富的人工智能和机器人研究人员，他最近致力于了解这些技术对组织和工作场所的影响，他提供了独特的视角。麦肯锡全球研究所是该领域研究的领导者，这次对话包含了许多关于正在展开的工作场所混乱性质的重要见解。

The individuals I spoke to offered a variety of viewpoints about this potential economic disruption and the type of policy solutions that might address it. In order to dive deeper into this topic, I turned to James Manyika, the Chairman of the McKinsey Global Institute. Manyika offers a unique perspective as an experienced AI and robotics researcher who has lately turned his efforts toward understanding the impact of these technologies on organizations and workplaces. The McKinsey Global Institute is a leader in conducting research into this area, and this conversation includes many important insights into the nature of the unfolding workplace disruption.

我向所有人提出的第二个问题涉及通往人类水平的人工智能（通常称为通用人工智能 (AGI)）的道路。从一开始，通用人工智能就是人工智能领域的圣杯。我想知道每个人对真正思考机器的前景、需要克服的障碍以及实现这一目标的时间框架有何看法。每个人都有重要的见解，但我发现三场对话特别有趣：Demis Hassabis 讨论了 DeepMind 正在进行的努力，这是规模最大、资金最充足的专门面向通用人工智能的计划。领导创建 IBM Watson 的团队的 David Ferrucci 现在是 Elemental Cognition 的首席执行官，这是一家希望通过利用对语言的理解来实现更通用智能的初创公司。Ray Kurzweil 现在负责 Google 的一个自然语言导向项目，他也对这个主题（以及许多其他主题）有重要的想法。Kurzweil 以其 2005 年出版的《奇点临近》一书而闻名。 2012年，他出版了一本关于机器智能的书《如何创造思维》，引起了拉里·佩奇的注意，并因此被谷歌聘用。

The second question I directed at everyone concerns the path toward human-level AI, or what is typically called Artificial General Intelligence (AGI). From the very beginning, AGI has been the holy grail of the field of artificial intelligence. I wanted to know what each person thought about the prospect for a true thinking machine, the hurdles that would need to be surmounted and the timeframe for when it might be achieved. Everyone had important insights, but I found three conversations to be especially interesting: Demis Hassabis discussed efforts underway at DeepMind, which is the largest and best funded initiative geared specifically toward AGI. David Ferrucci, who led the team that created IBM Watson, is now the CEO of Elemental Cognition, a startup that hopes to achieve more general intelligence by leveraging an understanding of language. Ray Kurzweil, who now directs a natural language-oriented project at Google, also had important ideas on this topic (as well as many others). Kurzweil is best known for his 2005 book, The Singularity is Near. In 2012, he published a book on machine intelligence, How to Create a Mind, which caught the attention of Larry Page and led to his employment at Google.

作为这些讨论的一部分，我看到了一个机会，可以请这群非常有成就的人工智能研究人员猜测一下 AGI 何时可能实现。我问的问题是：“您认为有 50% 的概率，哪一年可以实现人类级别的人工智能？”大多数参与者都愿意匿名提供他们的猜测。我在本书末尾的一个部分总结了这项非常非正式的调查的结果。有两个人愿意公开猜测，这将让您预览各种观点。雷·库兹韦尔认为，正如他之前多次表示的那样，人类级别的人工智能将在 2029 年左右实现——或者从撰写本文开始仅 11 年。另一方面，罗德尼·布鲁克斯猜测是 2200 年，即未来 180 多年。可以说，这里报道的对话中最有趣的方面之一是对一系列重要话题的截然不同的看法。

As part of these discussions, I saw an opportunity to ask this group of extraordinarily accomplished AI researchers to give me a guess for just when AGI might be realized. The question I asked was, “What year do you think human-level AI might be achieved, with a 50 percent probability?” Most of the participants preferred to provide their guesses anonymously. I have summarized the results of this very informal survey in a section at the end of this book. Two people were willing to guess on the record, and these will give you a preview of the wide range of opinions. Ray Kurzweil believes, as he has stated many times previously, that human-level AI will be achieved around 2029—or just eleven years from the time of this writing. Rodney Brooks, on the other hand, guessed the year 2200, or more than 180 years in the future. Suffice it to say that one of the most fascinating aspects of the conversations reported here is the starkly differing views on a wide range of important topics.

第三个讨论领域涉及人工智能发展在近期和更长远时期内将伴随的各种风险。一个已经变得明显的威胁是互联自主系统容易受到网络攻击或黑客攻击。随着人工智能越来越融入我们的经济和社会，解决这个问题将是我们面临的最关键挑战之一。另一个迫在眉睫的问题是机器学习算法容易产生偏见，在某些情况下，偏见是基于种族或性别的。我采访的许多人都强调了解决这个问题的重要性，并讲述了目前正在进行的该领域研究。一些人也表达了乐观的看法——暗示人工智能有朝一日可能会成为帮助对抗系统性偏见或歧视的有力工具。

The third area of discussion involves the varied risks that will accompany progress in artificial intelligence in both the immediate future and over much longer time horizons. One threat that is already becoming evident is the vulnerability of interconnected, autonomous systems to cyber attack or hacking. As AI becomes ever more integrated into our economy and society, solving this problem will be one of the most critical challenges we face. Another immediate concern is the susceptibility of machine learning algorithms to bias, in some cases on the basis of race or gender. Many of the individuals I spoke with emphasized the importance of addressing this issue and told of research currently underway in this area. Several also sounded an optimistic note—suggesting that AI may someday prove to be a powerful tool to help combat systemic bias or discrimination.

许多研究人员都对完全自主武器的威胁感到担忧。人工智能界的许多人认为，人工智能机器人或无人机具备杀人能力，无需人类“在场”授权采取任何致命行动，最终可能会像生物武器或化学武器一样危险和不稳定。2018 年 7 月，来自全球的 160 多家人工智能公司和 2,400 名研究人员（包括本文采访的许多人）签署了一份公开承诺，承诺永远不会开发此类武器。（https://futureoflife.org/lethal-autonomous-weapons-pledge/）本书中的几段对话深入探讨了武器化人工智能带来的危险。

A danger that many researchers are passionate about is the specter of fully autonomous weapons. Many people in the artificial intelligence community believe that AI-enabled robots or drones with the capability to kill, without a human “in the loop” to authorize any lethal action, could eventually be as dangerous and destabilizing as biological or chemical weapons. In July 2018, over 160 AI companies and 2,400 individual researchers from across the globe—including a number of the people interviewed here—signed an open pledge promising to never develop such weapons. (https://futureoflife.org/lethal-autonomous-weapons-pledge/) Several of the conversations in this book delve into the dangers presented by weaponized AI.

一个更具未来性和推测性的危险是所谓的“人工智能协调问题”。这是对真正智能或超级智能的机器可能会摆脱我们的控制，或者做出对人类产生不利影响的决定的担忧。正是这种恐惧引发了伊隆·马斯克等人的夸张言论。几乎每个与我交谈过的人都对这个问题发表了看法。为了确保我对这一担忧进行充分和平衡的报道，我采访了牛津大学人类未来研究所的尼克·博斯特罗姆。博斯特罗姆是畅销书《超级智能：道路、危险、策略》的作者，该书对可能比人类聪明得多的机器所带来的潜在风险进行了谨慎的论证。

A much more futuristic and speculative danger is the so-called “AI alignment problem.” This is the concern that a truly intelligent, or perhaps superintelligent, machine might escape our control, or make decisions that might have adverse consequences for humanity. This is the fear that elicits seemingly over-the-top statements from people like Elon Musk. Nearly everyone I spoke to weighed in on this issue. To ensure that I gave this concern adequate and balanced coverage, I spoke with Nick Bostrom of the Future of Humanity Institute at the University of Oxford. Bostrom is the author of the bestselling book Superintelligence: Paths, Dangers, Strategies, which makes a careful argument regarding the potential risks associated with machines that might be far smarter than any human being.

这里收录的对话是在 2018 年 2 月至 8 月期间进行的，几乎所有对话都持续了至少一个小时，有些甚至更长。这些对话都经过了录音、专业转录，然后由 Packt 团队进行了编辑，以使其更加清晰。最后，编辑后的文本被提供给与我交谈的人，他们随后有机会对其进行修改和扩展。因此，我完全有信心，这里记录的文字准确反映了我采访的人的想法。

The conversations included here were conducted from February to August 2018 and virtually all of them occupied at least an hour, some substantially more. They were recorded, professionally transcribed, and then edited for clarity by the team at Packt. Finally, the edited text was provided to the person I spoke to, who then had the opportunity to revise it and expand it. Therefore, I have every confidence that the words recorded here accurately reflect the thoughts of the person I interviewed.

我采访过的人工智能专家在国籍、所在地和所属关系方面差异很大。只要简单读一读这本书，你就会发现谷歌在人工智能社区中拥有巨大的影响力。在我采访的 23 人中，有 7 人目前或曾经与谷歌或其母公司 Alphabet 有关系。其他人才主要集中在麻省理工学院和斯坦福大学。杰夫·辛顿和约书亚·本吉奥分别就职于多伦多大学和蒙特利尔大学，加拿大政府利用其研究机构的声誉，将战略重点放在深度学习上。在我采访的 23 人中，有 19 人在美国工作。然而，在这 19 人中，超过一半出生在美国以外。他们的原籍国包括澳大利亚、中国、埃及、法国、以色列、罗得西亚（现为津巴布韦）、罗马尼亚和英国。我想说，这是技术移民在美国技术领导地位中发挥关键作用的相当有力的证据。

The AI experts I spoke to are highly varied in terms of their origins, locations, and affiliations. One thing that even a brief perusal of this book will make apparent is the outsized influence of Google in the AI community. Of the 23 people I interviewed, seven have current or former affiliations with Google or its parent, Alphabet. Other major concentrations of talent are found at MIT and Stanford. Geoff Hinton and Yoshua Bengio are based at the Universities of Toronto and Montreal respectively, and the Canadian government has leveraged the reputations of their research organizations into a strategic focus on deep learning. Nineteen of the 23 people I spoke to work in the United States. Of those 19, however, more than half were born outside the US. Countries of origin include Australia, China, Egypt, France, Israel, Rhodesia (now Zimbabwe), Romania, and the UK. I would say this is pretty dramatic evidence of the critical role that skilled immigration plays in the technological leadership of the US.

当我在本书中进行对话时，我脑海中浮现出各种各样的潜在读者，从专业的计算机科学家到管理人员和投资者，再到几乎任何对人工智能及其对社会的影响感兴趣的人。然而，一个特别重要的受众是年轻人，他们可能会考虑未来从事人工智能职业。目前，该领域的人才严重短缺，尤其是那些拥有深度学习技能的人，而人工智能或机器学习领域的职业将是令人兴奋、有利可图且意义重大的。

As I carried out the conversations in this book, I had in mind a variety of potential readers, ranging from professional computer scientists, to managers and investors, to virtually anyone with an interest in AI and its impact on society. One especially important audience, however, consists of young people who might consider a future career in artificial intelligence. There is currently a massive shortage of talent in the field, especially among those with skills in deep learning, and a career in AI or machine learning promises to be exciting, lucrative and consequential.

随着该行业努力吸引更多人才进入该领域，人们普遍认识到，必须采取更多措施来确保这些新人才更加多元化。如果人工智能确实有望重塑我们的世界，那么最了解这项技术并因此最有能力影响其发展方向的个人必须代表整个社会。

As the industry works to attract more talent into the field, there is widespread recognition that much more must be done to ensure that those new people are more diverse. If artificial intelligence is indeed poised to reshape our world, then it is crucial that the individuals who best understand the technology—and are therefore best positioned to influence its direction—be representative of society as a whole.

本书中接受采访的人中约有四分之一是女性，这一数字可能远高于整个人工智能或机器学习领域的女性。最近的一项研究发现，女性约占机器学习领域领先研究人员的 12%。( https://www.wired.com/story/artificial-intelligence-researchers-gender-imbalance ) 我采访的许多人都强调，需要增加女性和少数群体成员的代表性。

About a quarter of those interviewed in this book are women, and that number is likely significantly higher than what would be found across the entire field of AI or machine learning. A recent study found that women represent about 12 percent of leading researchers in machine learning. (https://www.wired.com/story/artificial-intelligence-researchers-gender-imbalance) A number of the people I spoke to emphasized the need for greater representation for both women and members of minority groups.

正如您将从本书对她的采访中了解到的那样，人工智能领域最杰出的女性之一对增加该领域的多样性有着特别的热情。斯坦福大学的李飞飞与他人共同创立了一个名为 AI4ALL ( http://ai-4-all.org/ ) 的组织，该组织专门为代表性不足的高中生提供以人工智能为重点的夏令营。AI4ALL 获得了业界的大力支持，包括最近来自谷歌的一笔资助，现在已扩大到包括美国六所大学的暑期项目。虽然还有许多工作要做，但有充分的理由乐观地认为，未来几年和几十年，人工智能研究人员的多样性将显著增加。

As you will learn from her interview in this book, one of the foremost women working in artificial intelligence is especially passionate about the need to increase diversity in the field. Stanford University’s Fei-Fei Li co-founded an organization now called AI4ALL (http://ai-4-all.org/) to provide AI-focused summer camps geared especially to underrepresented high school students. AI4ALL has received significant industry support, including a recent grant from Google, and has now scaled up to include summer programs at six universities across the United States. While much work remains to be done, there are good reasons to be optimistic that diversity among AI researchers will increase significantly in the coming years and decades.

虽然本书不需要技术背景，但您将会遇到与该领域相关的一些概念和术语。对于那些以前没有接触过人工智能的人来说，我相信这将提供一个机会，直接从该领域的一些顶尖人物那里了解这项技术。为了帮助经验不足的读者入门，本介绍之后简要概述了人工智能的词汇，我建议您在开始采访之前花点时间阅读这些材料。此外，对领先的人工智能教科书的合著者 Stuart Russell 的采访包括对该领域许多最重要思想的解释。

While this book does not assume a technical background, you will encounter some of the concepts and terminology associated with the field. For those without previous exposure to AI, I believe this will afford an opportunity to learn about the technology directly from some of the foremost minds in the field. To help less experienced readers get started, a brief overview of the vocabulary of AI follows this introduction, and I recommend you take a few moments to read this material before beginning the interviews. Additionally, the interview with Stuart Russell, who is the co-author of the leading AI textbook, includes an explanation of many of the field’s most important ideas.

我很荣幸能够参与本书的讨论。我相信你会发现，与我交谈的每一个人都深思熟虑、善于表达，并坚定地致力于确保他或她正在努力创造的技术能够造福人类。你很少会发现广泛的共识。本书充满了各种各样、往往相互尖锐冲突的见解、观点和预测。信息应该很明确：人工智能是一个广阔的领域。未来创新的性质、创新发生的速度以及它们将应用于的具体应用都笼罩在深深的不确定性之中。正是这种巨大的潜在颠覆性与根本性的不确定性相结合，使得我们必须开始就人工智能的未来及其对我们的生活方式可能意味着什么进行有意义且包容的对话。我希望这本书能为这一讨论做出贡献。

It has been an extraordinary privilege for me to participate in the conversations in this book. I believe you will find everyone I spoke with to be thoughtful, articulate, and deeply committed to ensuring that the technology he or she is working to create will be leveraged for the benefit of humanity. What you will not so often find is broad-based consensus. This book is full of varied, and often sharply conflicting, insights, opinions, and predictions. The message should be clear: Artificial intelligence is a wide open field. The nature of the innovations that lie ahead, the rate at which they will occur, and the specific applications to which they will be applied are all shrouded in deep uncertainty. It is this combination of massive potential disruption together with fundamental uncertainty that makes it imperative that we begin to engage in a meaningful and inclusive conversation about the future of artificial intelligence and what it may mean for our way of life. I hope this book will make a contribution to that discussion.

蒙特利尔学习算法研究所科学主任、蒙特利尔大学计算机科学与运筹学教授

SCIENTIFIC DIRECTOR, MONTREAL INSTITUTE FOR LEARNING ALGORITHMS AND PROFESSOR OF COMPUTER SCIENCE AND OPERATIONS RESEARCH, UNIVERSITY OF MONTREAL

约书亚·本吉奥是蒙特利尔大学计算机科学与运筹学教授，被公认为深度学习的先驱之一。约书亚在推动神经网络研究方面发挥了重要作用，尤其是“无监督”学习，即神经网络可以在不依赖大量训练数据的情况下进行学习。

Yoshua Bengio is a professor of computer science and operations research at the University of Montreal and is widely recognized as one of the pioneers of deep learning. Yoshua was instrumental in advancing neural network research, in particular “unsupervised” learning where neural networks can learn without relying on vast amounts of training data.

马丁·福特：您处于人工智能研究的前沿，所以我首先想问一下，您认为未来几年我们会在哪些当前研究问题上取得突破，以及这些问题将如何帮助我们走上 AGI（通用人工智能）的道路？

MARTIN FORD: You are at the forefront of AI research, so I want to begin by asking what current research problems you think we’ll see breakthroughs in over the next few years, and how those will help us on the road to AGI (artificial general intelligence)?

YOSHUA BENGIO：我不知道我们到底会看到什么，但我可以告诉你，我们面前确实存在一些非常困难的问题，而且我们距离人类水平的人工智能还很遥远。研究人员正在试图了解问题所在，比如，为什么我们不能制造出像我们一样真正理解世界的机器？是因为我们没有足够的训练数据，还是我们没有足够的计算能力？我们中的许多人认为，我们还缺少所需的基本要素，例如理解数据中因果关系的能力——这种能力实际上使我们能够概括并在与我们接受过训练的环境截然不同的环境中得出正确的答案。

YOSHUA BENGIO: I don’t know exactly what we’re going to see, but I can tell you that there are some really hard problems in front of us and that we are far from human-level AI. Researchers are trying to understand what the issues are, such as, why is it that we can’t build machines that really understand the world as well as we do? Is it just that we don’t have enough training data, or is it that we don’t have enough computing power? Many of us think that we are also missing the basic ingredients needed, such as the ability to understand causal relationships in data—an ability that actually enables us to generalize and to come up with the right answers in settings that are very different from those we’ve been trained in.

人类可以想象自己经历一种对他们来说完全陌生的经历。例如，你可能从未经历过车祸，但你可以想象一场车祸，而且由于你已经知道的所有事情，你实际上能够扮演角色并做出正确的决定，至少在你的脑海中是这样。当前的机器学习基于监督学习，计算机本质上是学习它所看到的数据的统计数据，并且需要手动完成该过程。换句话说，人类必须提供所有这些标签，可能是数亿个正确答案，然后计算机可以从中学习。

A human can imagine themselves going through an experience that is completely new to them. You might have never had a car accident, for example, but you can imagine one and because of all the things you already know you’re actually able to roleplay and make the right decisions, at least in your head. Current machine learning is based on supervised learning, where a computer essentially learns about the statistics of the data that it sees, and it needs to be taken through that process by hand. In other words, humans have to provide all of those labels, possibly hundreds of millions of correct answers, that the computer can then learn from.

目前很多研究都是在我们做得不太好的领域，比如无监督学习。计算机可以更自主地获取有关世界的知识。另一个研究领域是因果关系，计算机不仅可以观察数据，比如图像或视频，还可以对其采取行动，观察这些行动的影响，从而推断出世界上的因果关系。例如，DeepMind、OpenAI 或伯克利正在用虚拟代理做的事情正朝着正确的方向发展，可以回答这类问题，我们也在蒙特利尔做这类事情。

A lot of current research is in areas where we’re not doing so well, such as unsupervised learning. This is where the computer can be more autonomous in the way that it acquires knowledge about the world. Another area of research is in causality, where the computer can not only observe data, like images or videos, but also act on it and see the effect of those actions in order to infer causal relationships in the world. The kinds of things that DeepMind, OpenAI, or Berkeley are doing with virtual agents, for example, are going in the right direction to answer those types of questions, and we’re also doing these kinds of things in Montreal.

马丁·福特：您认为目前有哪些项目真正处于深度学习的前沿？最明显的是 AlphaZero，但还有哪些项目真正代表了这项技术的前沿？

MARTIN FORD: Are there any particular projects that you would point to as being really at the forefront of deep learning right now? The obvious one is AlphaZero, but what other projects really represent the leading edge of this technology?

YOSHUA BENGIO：有很多有趣的项目，但我认为从长远来看可能产生重大影响的是那些涉及虚拟世界的项目，在虚拟世界中，代理会尝试解决问题并尝试了解其环境。我们正在 MILA 进行这项工作，DeepMind、OpenAI、伯克利、Facebook 和 Google Brain 也正在开展同一领域的项目。这是新的前沿。

YOSHUA BENGIO: There are a number of interesting projects, but the ones that I think are likely in the long run to have a big impact are those that involve virtual worlds in which an agent is trying to solve problems and is trying to learn about their environment. We are working on this at MILA, and there are projects in the same area in progress at DeepMind, OpenAI, Berkeley, Facebook and Google Brain. It’s the new frontier.

但需要记住的是，这不是短期研究。我们不是研究深度学习的某个特定应用，而是研究学习代理如何理解其环境，以及学习代理如何学习说话或理解语言，特别是我们所说的扎实语言。

It’s important to remember, though, that this is not short-term research. We’re not working on a particular application of deep learning, instead we’re looking into the future of how a learning agent makes sense of its environment and how a learning agent can learn to speak or to understand language, in particular what we call grounded language.

马丁·福特：您能解释一下这个术语吗？

MARTIN FORD: Can you explain that term?

YOSHUA BENGIO：当然，之前为了让计算机理解语言所做的大量努力都是让计算机阅读大量文本。这很好，但除非这些句子与真实事物相关，否则计算机很难真正理解这些单词的含义。例如，您可以将单词链接到图像或视频，或者链接到可能是现实世界中的物体的机器人。

YOSHUA BENGIO: Sure, a lot of the previous effort in trying to make computers understand language has the computer just read lots and lots of text. That’s nice and all, but it’s hard for the computer to actually get the meaning of those words unless those sentences are associated with real things. You might link words to images or videos, for example, or for robots that might be objects in the real world.

目前，有大量扎实的语言学习研究试图建立对语言的理解，即使它只是语言的一小部分，计算机也能真正理解这些单词的含义，并能根据这些单词做出相应的反应。这是一个非常有趣的方向，可能会对对话语言理解、个人助理等产生实际影响。

There’s a lot of research in grounded language learning now trying to build an understanding of language, even if it’s a small subset of the language, where the computer actually understands what those words mean, and it can act in correspondence to those words. It’s a very interesting direction that could have a practical impact on things like language understanding for dialog, personal assistants, and so on.

马丁·福特：那么，这个想法基本上就是让代理在模拟环境中自由活动，并让它像小孩一样学习？

MARTIN FORD: So, the idea there is basically to turn an agent loose in a simulated environment and have it learn like a child?

YOSHUA BENGIO：没错，事实上，我们希望从儿童发展科学家那里获得灵感，他们正在研究新生儿在生命最初几个月内如何经历一系列阶段，逐渐对世界有更多的了解。我们并不完全了解其中哪些部分是天生的，哪些是后天习得的，我认为了解婴儿经历的过程可以帮助我们设计自己的系统。

YOSHUA BENGIO: Exactly, in fact, we want to take inspiration from child development scientists who are studying how a newborn goes through a series of stages in the first few months of life where they gradually acquire more understanding about the world. We don’t completely understand which part of this is innate or really learned, and I think this understanding of what babies go through can help us design our own systems.

几年前，我在机器学习中引入了一个在训练动物时非常常见的理念，那就是课程学习。这个理念就是，我们不会将所有训练示例以任意顺序显示为一大堆。相反，我们会按照学习者能够理解的顺序来学习示例。我们从简单的东西开始，一旦掌握了简单的东西，我们就可以使用这些概念作为学习稍微复杂一点的东西的基础。这就是我们上学的原因，也是为什么我们 6 岁时不会直接上大学的原因。这种学习在训练计算机方面也变得越来越重要。

One idea I introduced a few years ago in machine learning that is very common in training animals is curriculum learning. The idea is that we don’t just show all the training examples as one big pile in an arbitrary order. Instead, we go through examples in an order that makes sense for the learner. We start with easy things, and once the easy things are mastered, we can use those concepts as the building blocks for learning slightly more complicated things. That’s why we go through school, and why when we are 6 years old we don’t go straight to university. This kind of learning is becoming more important in training computers as well.

马丁·福特：我们来谈谈 AGI 之路。显然，您认为无监督学习（本质上是让系统像人一样学习）是 AGI 的重要组成部分。这足以实现 AGI 吗？还是说，要实现 AGI，还需要其他关键组成部分和突破？

MARTIN FORD: Let’s talk about the path to AGI. Obviously, you believe that unsupervised learning—essentially having a system learn like a person—is an important component of it. Is that enough to get to AGI, or are there other critical components and breakthroughs that have to happen for us to get there?

YOSHUA BENGIO：我的朋友 Yann LeCun 有一个很好的比喻来描述这一点。我们现在正在爬山，我们都很兴奋，因为我们在爬山方面取得了很大进展，但当我们接近山顶时，我们开始看到前面又出现了一系列其他的山丘。这就是我们现在在 AGI 的发展中所看到的，也是我们当前方法的一些局限性。当我们爬第一座山时，比如当我们在探索如何训练更深层次的网络时，我们没有看到我们正在构建的系统的局限性，因为我们只是在探索如何往上走几步。

YOSHUA BENGIO: My friend Yann LeCun has a nice metaphor that describes this. We’re currently climbing a hill, and we are all excited because we have made a lot of progress on climbing the hill, but as we approach the top of the hill, we can start to see a series of other hills rising in front of us. That is what we see now in the development of AGI, some of the limitations of our current approaches. When we were climbing the first hill, when we were discovering how to train deeper networks, for example, we didn’t see the limitations of the systems we were building because we were just discovering how to go up a few steps.

当我们的技术取得令人满意的进步时——我们登上了第一座山的顶峰——我们也看到了局限性，然后我们看到了另一座我们必须攀登的山，一旦我们爬上了那座山，我们就会再看到下一座山，依此类推。我们无法预知在达到人类智能水平之前还需要多少突破或重大进步。

As we reach this satisfying improvement that we are getting in our techniques—we reach the top of the first hill—we also see the limitations, and then we see another hill that we have to climb, and once we climb that one we’ll see another one, and so on. It’s impossible to tell how many more breakthroughs or significant advances are going to be needed before we reach human-level intelligence.

马丁·福特：有多少座山？AGI 的时间表是什么？你能给我一个最佳的猜测吗？

MARTIN FORD: How many hills are there? What’s the timescale for AGI? Can you give me your best guess?

约书亚·本吉奥：你不会从我这里得到这些，没有意义。猜测日期是没有用的，因为我们没有线索。我只能说，这不会在未来几年内发生。

YOSHUA BENGIO: You won’t be getting that from me, there’s no point. It’s useless to guess a date because we have no clue. All I can say is that it’s not going to happen in the next few years.

马丁·福特：您认为深度学习或神经网络真的是未来的发展方向吗？

MARTIN FORD: Do you think that deep learning or neural networks generally are really the way forward?

YOSHUA BENGIO：是的，就深度学习背后的科学概念以及该领域多年来取得的进展而言，我们发现，在很大程度上，深度学习和神经网络背后的许多概念都将继续存在。简而言之，它们非常强大。事实上，它们可能会帮助我们更好地理解动物和人类大脑如何学习复杂的东西。但正如我所说，它们还不足以让我们实现 AGI。我们目前所处的阶段可以看到我们目前所拥有的一些局限性，我们将在此基础上进行改进和发展。

YOSHUA BENGIO: Yes, what we have discovered in terms of the scientific concepts that are behind deep learning and the years of progress made in this field, means that for the most part, many of the concepts behind deep learning and neural networks are here to stay. Simply put, they are incredibly powerful. In fact, they are probably going to help us better understand how animal and human brains learn complex things. As I said, though, they’re not enough to get us to AGI. We’re at a point where we can see some of the limitations in what we currently have, and we’re going to improve and build on top of that.

马丁·福特：我知道艾伦人工智能研究所正在开展“马赛克计划”，旨在将常识植入计算机。您认为这种事情至关重要吗？或者您认为常识可能是学习过程的一部分？

MARTIN FORD: I know that the Allen Institute for AI is working on Project Mosaic, which is about building common sense into computers. Do you think that kind of thing is critical, or do you think that maybe common sense emerges as part of the learning process?

YOSHUA BENGIO：我相信常识会作为学习过程的一部分出现。它不会因为有人把一些知识塞进你的脑袋里而出现，人类的运作方式不是这样的。

YOSHUA BENGIO: I’m sure common sense will emerge as part of the learning process. It won’t come up because somebody sticks little bits of knowledge into your head, that’s not how it works for humans.

马丁·福特：深度学习是实现 AGI 的主要方式吗？或者您认为它需要某种混合系统？

MARTIN FORD: Is deep learning the primary way to get us to AGI, or do you think it’s going to require some sort of a hybrid system?

YOSHUA BENGIO：传统人工智能纯粹是符号化的，没有学习。它专注于认知的一个非常有趣的方面，即我们如何按顺序推理和组合信息。另一方面，深度学习神经网络一直专注于一种自下而上的认知视角，我们从感知开始，将机器对世界的理解锚定在感知中。从那里，我们构建分布式表示并可以捕捉许多变量之间的关系。

YOSHUA BENGIO: Classical AI was purely symbolic, and there was no learning. It focused on a really interesting aspect of cognition, which is how we sequentially reason and combine pieces of information. Deep learning neural networks, on the other hand, have always been about focusing on a sort of bottom-up view of cognition, where we start with perception and we anchor the machine’s understanding of the world in perception. From there, we build distributed representations and can capture the relationship between many variables.

1999 年左右，我和弟弟研究了这些变量之间的关系。这引发了自然语言领域的许多最新进展，例如词嵌入，即单词和句子的分布式表示。在这些情况下，一个单词由大脑中的活动模式或一组数字表示。那些具有相似含义的单词会与相似的数字模式相关联。

I studied the relationships between such variables with my brother around 1999. That gave rise to a lot of the recent progress in natural language, such as word embeddings, or distributed representations for words and sentences. In these cases, a word is represented by a pattern of activity in your brain—or by a set of numbers. Those words that have a similar meaning are then associated with similar patterns of numbers.

目前深度学习领域的进展是，人们在这些深度学习概念的基础上开始尝试解决经典人工智能的推理和理解、编程或规划问题。研究人员正试图利用我们从感知中开发出的构建模块，并将其扩展到这些更高级的认知任务（有时被心理学家称为系统 2）。我相信，在某种程度上，这是我们走向人类水平人工智能的方式。这并不是说它是一个混合系统；就像我们试图解决一些经典人工智能试图解决的相同问题，但使用来自深度学习的构建模块。这是一种非常不同的做法，但目标非常相似。

What’s going on now in the deep learning field is that people are building on top of these deep learning concepts and starting to try to solve the classical AI problems of reasoning and being able to understand, program, or plan. Researchers are trying to use the building blocks that we developed from perception and extend them towards these higher-level cognitive tasks (sometimes called System 2 by psychologists). I believe in part that’s the way that we’re going to move towards human-level AI. It’s not that it’s a hybrid system; it’s like we’re trying to solve some of the same problems that classical AI was trying to solve but using the building blocks coming from deep learning. It’s a very different way of doing it, but the objectives are very similar.

马丁·福特：那么您预测它们都将是神经网络，只是架构有所不同？

MARTIN FORD: Your prediction, then, is that it’s all going to be neural networks, but with different architectures?

YOSHUA BENGIO：是的。请注意，你的大脑全是神经网络。我们必须想出不同的架构和不同的训练框架，它们可以完成传统人工智能试图做的事情，比如推理、推断你所看到和计划的事情的解释。

YOSHUA BENGIO: Yes. Note that your brain is all neural networks. We have to come up with different architectures and different training frameworks that can do the kinds of things that classical AI was trying to do, like reasoning, inferring an explanation for what you’re seeing and planning.

马丁·福特：您认为这一切都可以通过学习和培训来实现吗，还是需要一定的结构？

MARTIN FORD: Do you think it can all be done with learning and training or does there need to be some structure there?

YOSHUA BENGIO：这里面有结构，只是它不是我们编写百科全书或数学公式时用来表示知识的那种结构。我们放入的结构对应于神经网络的架构，也对应于对世界和我们试图解决的任务的相当广泛的假设。当我们放入一个特殊的结构和架构，让网络拥有注意力机制时，它就放入了大量的先验知识。事实证明，这是机器翻译等技术成功的关键。

YOSHUA BENGIO: There is structure there, it’s just that it’s not the kind of structure that we use to represent knowledge when we write an encyclopedia, or we write a mathematical formula. The kind of structure that we put in corresponds to the architecture of the neural net, and to fairly broad assumptions about the world and the kind of task that we’re trying to solve. When we put in a special structure and architecture that allows the network to have an attention mechanism, it’s putting in a lot of prior knowledge. It turns out that this is central to the success of things like machine translation.

你需要在工具箱中拥有这种工具来解决其中的一些问题，就像处理图像时，你需要拥有类似卷积神经网络结构的东西才能做好工作一样。如果你不加入这种结构，那么性能就会差很多。关于世界和你试图学习的功能，已经有很多领域特定的假设，这些假设隐含在深度学习中使用的架构和训练目标中。这就是当今大多数研究论文的主题。

You need that kind of tool in your toolbox in order to solve some of those problems, in the same way that if you deal with images, you need to have something like a convolutional neural network structure in order to do a good job. If you don’t put in that structure, then performance is much worse. There are already a lot of domain-specific assumptions about the world and about the function you’re trying to learn, that are implicit in the kind of architectures and training objectives that are used in deep learning. This is what most of the research papers today are about.

马丁·福特：我提出结构问题的目的是，比如，婴儿一出生就能识别人脸。显然，人类大脑中存在某种结构，让婴儿能够做到这一点。这不仅仅是处理像素的原始神经元。

MARTIN FORD: What I was trying to get at with the question on structure was that, for example, a baby can recognize human faces right after it is born. Clearly, then, there is some structure in the human brain that allows the baby to do that. It’s not just raw neurons working on pixels.

YOSHUA BENGIO：你错了！这是原始神经元在像素上工作，只不过婴儿的大脑中存在一种特殊的结构，可以识别里面有两个点的圆形物体。

YOSHUA BENGIO: You’re wrong! It is raw neurons working on pixels, except that there is a particular architecture in the baby’s brain that recognizes something circular with two dots inside it.

马丁·福特：我的观点是，该结构是预先存在的。

MARTIN FORD: My point is that the structure pre-exists.

YOSHUA BENGIO：当然有，但是我们在这些神经网络中设计的所有东西也都是预先存在的。深度学习研究人员所做的工作就像进化的工作，我们以架构和训练程序的形式将先验知识放入其中。

YOSHUA BENGIO: Of course it does, but all the things that we’re designing in those neural networks also pre-exist. What deep learning researchers are doing is like the work of evolution, where we’re putting in the prior knowledge in the form of both the architecture and the training procedure.

如果我们愿意，我们可以硬连线一些东西，让网络能够识别人脸，但这对人工智能来说毫无用处，因为它们可以很快学会。相反，我们会加入真正有助于解决我们试图解决的难题的东西。

If we wanted, we could hardwire something that would allow the network to recognize a face, but it’s useless for an AI because they can learn that very quickly. Instead, we put in the things that are really useful for solving the harder problems that we’re trying to deal with.

没有人说人类、婴儿和动物没有先天知识，事实上，大多数动物只有先天知识。蚂蚁学得不多，它们就像一个庞大的固定程序，但随着智力等级的提高，学习的比例不断增加。人类与许多其他动物的不同之处在于，我们学习了多少，而一开始先天又有多少。

Nobody is saying that there is no innate knowledge in humans, babies, and animals, in fact, most animals have only innate knowledge. An ant doesn’t learn much, it’s all like a big, fixed program, but as you go higher up in the intelligence hierarchy, the share of learning keeps increasing. What makes humans different from many other animals is how much we learn versus how much is innate at the start.

马丁·福特：让我们回顾一下，定义一些概念。在 20 世纪 80 年代，神经网络是一个非常边缘化的学科，它们只有一层，所以没有什么深度。你参与了将其转变为我们现在所说的深度学习。你能用相对非技术性的术语来定义它是什么吗？

MARTIN FORD: Let’s step back and define some of those concepts. In the 1980s, neural networks were a very marginalized subject and they were just one layer, so there was nothing deep about them. You were involved in transforming that into what we now call deep learning. Could you define, in relatively non-technical terms, what that is?

YOSHUA BENGIO：深度学习是机器学习的一种方法。机器学习试图通过让计算机从例子中学习来将知识输入计算机，而深度学习则是以一种受大脑启发的方式来进行。

YOSHUA BENGIO: Deep learning is an approach to machine learning. While machine learning is trying to put knowledge into computers by allowing computers to learn from examples, deep learning is doing it in a way that is inspired by the brain.

深度学习和机器学习只是早期神经网络研究的延续。它们之所以被称为“深度”，是因为它们增加了训练更深层网络的能力，这意味着它们有更多的层，每一层代表不同的表示级别。我们希望随着网络变得越来越深，它可以表示更抽象的东西，到目前为止，情况似乎确实如此。

Deep learning and machine learning are just a continuation of that earlier work on neural networks. They’re called “deep” because they added the ability to train deeper networks, meaning they have more layers, and each layer represents a different level of representation. We hope that as the network gets deeper, it can represent more abstract things, and so far, that does seem to be the case.

马丁·福特：您说的层次是指抽象层次吗？那么，就视觉图像而言，第一层是像素，然后是边缘，接着是角落，然后逐渐达到物体？

MARTIN FORD: When you say layers, do you mean layers of abstraction? So, in terms of a visual image, the first layer would be pixels, then it would be edges, followed by corners, and then gradually you would get all the way up to objects?

约书亚·本吉奥：是的，正确。

YOSHUA BENGIO: Yes, that’s correct.

马丁·福特：如果我理解正确的话，计算机仍然不明白那个物体是什么，对吗？

MARTIN FORD: If I understand correctly, though, the computer still doesn’t understand what that object is, right?

YOSHUA BENGIO：计算机有一定的理解能力，这不是一个非黑即白的争论。猫能理解门，但它理解得不如你。不同的人对周围许多事物的理解程度不同，而科学就是试图加深我们对这些事物的理解。如果这些网络接受过图像训练，它们对图像有一定的理解能力，但这种理解水平仍然不像我们的那么抽象和笼统。其中一个原因是，我们在对世界的三维理解的背景下解释图像，这种理解得益于我们的立体视觉以及我们在世界上的运动和行为。这不仅为我们提供了视觉模型，还为我们提供了物体的物理模型。目前计算机对图像的理解水平仍然很原始，但已经足够好，在许多应用中非常有用。

YOSHUA BENGIO: The computer has some understanding, it’s not a black-and-white argument. A cat understands a door, but it doesn’t understand it as well as you do. Different people have different levels of understanding of the many things around them, and science is about trying to deepen our understanding of those many things. These networks have a level of understanding of images if they’ve been trained on images, but that level is still not as abstract and as general as ours. One reason for this is that we interpret images in the context of our three-dimensional understanding of the world, obtained thanks to our stereo vision and our movements and actions in the world. This gives us a lot more than just a visual model: it also gives us a physical model of objects. The current level of computer understanding of images is still primitive but it’s still good enough to be incredibly useful in many applications.

马丁·福特：真正让深度学习成为可能的是反向传播，这是真的吗？反向传播的理念是，你可以通过层将错误信息发送回去，并根据最终结果调整每一层。

MARTIN FORD: Is it true that the thing that has really made deep learning possible is backpropagation? The idea that you can send the error information back through the layers, and adjust each layer based on the final outcome.

YOSHUA BENGIO：事实上，反向传播是近年来深度学习成功的核心。它是一种进行信用分配的方法，即弄清楚内部神经元应如何变化才能使更大的网络正常运行。反向传播，至少在神经网络的背景下，是在 20 世纪 80 年代初发现的，当时我开始了自己的工作。Yann LeCun 与 Geoffrey Hinton 和 David Rumelhart 大约在同一时间独立发现了它。这是一个古老的想法，但直到 2006 年左右，也就是四分之一世纪之后，我们才在实际训练这些更深层次的网络方面取得成功。

YOSHUA BENGIO: Indeed, backpropagation has been at the heart of the success of deep learning in recent years. It is a method to do credit assignment, that is, to figure out how internal neurons should change to make the bigger network behave properly. Backpropagation, at least in the context of neural networks, was discovered in the early 1980s, at the time when I started my own work. Yann LeCun independently discovered it around the same time as Geoffrey Hinton and David Rumelhart. It’s an old idea, but we didn’t practically succeed in training these deeper networks until around 2006, over a quarter of a century later.

从那时起，我们就为这些网络添加了许多其他功能，这对于我们的人工智能研究非常令人兴奋，例如注意力机制、记忆力，以及不仅可以分类还可以生成图像的能力。

Since then, we’ve been adding a number of other features to these networks, which are very exciting for our research into artificial intelligence, such as attention mechanisms, memory, and the ability to not just classify but also generate images.

马丁·福特：我们知道大脑是否会进行类似反向传播的事情吗？

MARTIN FORD: Do we know if the brain does something similar to backpropagation?

YOSHUA BENGIO：这个问题问得很好。神经网络并不是试图模仿大脑，而是至少在抽象层面上受到了大脑某些计算特性的启发。

YOSHUA BENGIO: That’s a good question. Neural nets are not trying to imitate the brain, but they are inspired by some of its computational characteristics, at least at an abstract level.

你必须意识到，我们还没有完全了解大脑是如何运作的。大脑的许多方面尚未被神经科学家所理解。关于大脑的观察有很多，但我们还不知道如何将这些点连接起来。

You have to realize that we don’t yet have a full picture of how the brain works. There are many aspects of the brain that are not yet understood by neuroscientists. There are tons of observations about the brain, but we don’t know how to connect the dots yet.

也许我们在机器学习中利用神经网络所做的工作可以为脑科学提供一个可检验的假设。这是我感兴趣的事情之一。特别是，到目前为止，反向传播大多被认为是计算机可以做的事情，但对大脑来说并不现实。

It may be that the work that we’re doing in machine learning with neural nets could provide a testable hypothesis for brain science. That’s one of the things that I’m interested in. In particular, backpropagation up to now has mostly been considered something that computers can do, but not realistic for brains.

事实上，反向传播的效果非常好，这表明大脑可能在做一些类似的事情——不完全一样，但功能相同。因此，我目前正在参与这个方向的一些非常有趣的研究。

The thing is, backpropagation is working incredibly well, and it suggests that maybe the brain is doing something similar—not exactly the same, but with the same function. As a result of that, I’m currently involved in some very interesting research in that direction.

马丁·福特：我知道曾经有过“人工智能寒冬”，当时大多数人都放弃了深度学习，但像你、杰弗里·辛顿和扬·勒昆这样的少数人却让它继续存在。后来它是如何演变到今天的地步的？

MARTIN FORD: I know that there was an “AI Winter” where most people had dismissed deep learning, but a handful of people, like yourself, Geoffrey Hinton, and Yann LeCun, kept it alive. How did that then evolve to the point where we find ourselves today?

YOSHUA BENGIO：在 90 年代末到 21 世纪初，神经网络并不流行，很少有团队参与其中。我有一种强烈的直觉，如果抛弃神经网络，我们就会抛弃一些非常重要的东西。

YOSHUA BENGIO: By the end of the ‘90s and through the early 2000s, neural networks were not trendy, and very few groups were involved with them. I had a strong intuition that by throwing out neural networks, we were throwing out something really important.

部分原因是我们现在称之为组合性的东西：这些系统能够以组合的方式表示有关数据的非常丰富的信息，其中你可以组合许多与神经元和层相对应的构建块。这让我想到了语言模型，即使用词嵌入对文本进行建模的早期神经网络。每个单词都与一组数字相关联，这些数字对应于由机器自主学习的不同属性。它当时并没有真正流行起来，但现在几乎所有与从数据建模语言有关的事情都使用这些想法。

Part of that was because of something that we now call compositionality: The ability of these systems to represent very rich information about the data in a compositional way, where you compose many building blocks that correspond to the neurons and the layers. That led me to language models, early neural networks that model text using word embeddings. Each word is associated with a set of numbers corresponding to different attributes that are learned autonomously by the machine. It didn’t really catch on at the time, but nowadays almost everything to do with modeling language from data uses these ideas.

最大的问题是如何训练更深的网络，而 Geoffrey Hinton 及其对受限玻尔兹曼机 (RBM) 的研究取得了突破。在我的实验室里，我们研究的是自动编码器，它与 RBM 密切相关，而自动编码器已经催生了各种模型，例如生成对抗网络。事实证明，通过堆叠这些 RBM 或自动编码器，我们能够训练比以前更深的网络。

The big question was how we could train deeper networks, and the breakthrough was made by Geoffrey Hinton and his work with Restricted Boltzmann Machines (RBMs). In my lab, we were working on autoencoders, which are very closely related to RBMs, and autoencoders have given rise to all kinds of models, such as generative adversarial networks. It turned out that by stacking these RBMs or autoencoders we are able to train deeper networks than we were able to before.

马丁福特：你能解释一下什么是自动编码器吗？

MARTIN FORD: Could you explain what an autoencoder is?

YOSHUA BENGIO：自动编码器由两个部分组成：编码器和解码器。其理念是，编码器部分获取图像，并尝试以压缩方式（例如口头描述）表示图像。然后，解码器获取该表示并尝试恢复原始图像。自动编码器经过训练可以进行这种压缩和解压缩，以便尽可能忠实于原始图像。

YOSHUA BENGIO: There are two parts to an autoencoder, an encoder and a decoder. The idea is that the encoder part takes an image, for example, and tries to represent it in a compressed way, such as a verbal description. The decoder then takes that representation and tries to recover the original image. The autoencoder is trained to do this compression and decompression so that it is as faithful as possible to the original.

自最初的设想以来，自动编码器已经发生了很大变化。现在，我们认为它们是将原始信息（如图像）转换为更抽象的空间，以便更容易读取其中重要的语义方面。这是编码器部分。解码器反向工作，获取那些高级数量（您不必手动定义）并将它们转换为图像。这是早期的深度学习工作。

Autoencoders have changed quite a bit since that original vision. Now, we think of them in terms of taking raw information, like an image, and transforming it into a more abstract space where the important, semantic aspect of it will be easier to read. That’s the encoder part. The decoder works backwards, taking those high-level quantities—that you don’t have to define by hand—and transforming them into an image. That was the early deep learning work.

几年后，我们发现训练深度网络并不需要这些方法，我们只需改变非线性即可。我的一名学生正在与神经科学家合作，我们认为应该尝试整流线性单元 (ReLU)——当时我们称之为整流器——因为它们在生物学上更合理，这是一个真正从大脑中汲取灵感的例子。

Then a few years later, we discovered that we didn’t need these approaches to train deep networks, we could just change the nonlinearity. One of my students was working with neuroscientists, and we thought that we should try rectified linear units (ReLUs)—we called them rectifiers in those days—because they were more biologically plausible, and this is an example of actually taking inspiration from the brain.

马丁·福特：你从这一切中学到了什么？

MARTIN FORD: What did you learn from all of that?

YOSHUA BENGIO：我们之前使用 S 型函数来训练神经网络，但事实证明，使用 ReLU 后，我们可以更轻松地训练非常深的网络。这是 2010 年或 2011 年左右发生的另一个重大变化。

YOSHUA BENGIO: We had previously used a sigmoid function to train neural nets, but it turned out that by using ReLUs we could suddenly train very deep nets much more easily. That was another big change that occurred around 2010 or 2011.

计算机视觉领域有一个非常大的数据集，即 ImageNet 数据集，只有当我们的深度学习方法能够在该数据集上取得良好结果时，该领域的人们才会相信我们的方法。Geoffrey Hinton 的团队确实做到了这一点，他们延续了 Yann LeCun 早期在卷积网络（即专门用于图像的神经网络）方面的工作。2012 年，这些经过额外改进的新型深度学习架构获得了巨大成功，并大大改进了现有方法。几年之内，整个计算机视觉社区都转向了这类网络。

There is a very large dataset—the ImageNet dataset—which is used in computer vision, and people in that field would only believe in our deep learning methods if we could show good results on that dataset. Geoffrey Hinton’s group actually did it, following up on earlier work by Yann LeCun on convolutional networks—that is, neural networks which were specialized for images. In 2012, these new deep learning architectures with extra twists were used with huge success and showed a big improvement on existing methods. Within a couple of years, the whole computer vision community switched to these kinds of networks.

马丁·福特：那么，这就是深度学习真正起飞的时刻吗？

MARTIN FORD: So that’s the point at which deep learning really took off?

YOSHUA BENGIO：时间稍晚一些。到 2014 年，深度学习在社区中的应用开始大幅加速。

YOSHUA BENGIO: It was a bit later. By 2014, things were lining up for a big acceleration in the community for the take-up of deep learning.

马丁·福特：那时候它就从以大学为中心转变为谷歌、Facebook和百度等主流领域了？

MARTIN FORD: That’s when it transitioned from being centered in universities to being in the mainstream domain at places like Google, Facebook, and Baidu?

YOSHUA BENGIO：没错。这种转变开始得稍早一些，大约在 2010 年，当时谷歌、IBM 和微软等公司正在研究用于语音识别的神经网络。到 2012 年，谷歌的 Android 智能手机上就安装了这些神经网络。这是革命性的，因为同一项深度学习技术可以用于计算机视觉和语音识别。它引起了人们对该领域的极大关注。

YOSHUA BENGIO: Exactly. The shift started slightly earlier, around 2010, with companies like Google, IBM, and Microsoft, who were working on neural networks for speech recognition. By 2012, Google had these neural networks on their Android smartphones. It was revolutionary for the fact that the same technology of deep learning could be used for both computer vision and speech recognition. It drove a lot of attention toward the field.

马丁·福特：回想您第一次开始研究神经网络的时候，您是否对它取得的进展以及它已成为谷歌和 Facebook 等大公司现在所做事情的核心感到惊讶？

MARTIN FORD: Thinking back to when you first started in neural networks, are you surprised at the distance things have come and the fact that they’ve become so central to what large companies, like Google and Facebook, are doing now?

YOSHUA BENGIO：当然，我们没想到这一点。我们在深度学习方面取得了一系列重要且令人惊讶的突破。我之前提到过，语音识别出现于 2010 年左右，然后计算机视觉出现于 2012 年左右。几年后，在 2014 年和 2015 年，我们在机器翻译方面取得了突破，这些突破最终在 2016 年应用于谷歌翻译。2016 年也是我们见证 AlphaGo 突破的一年。所有这些，以及其他一些事情，都是我们真的没有预料到的。

YOSHUA BENGIO: Of course, we didn’t expect that. We’ve had a series of important and surprising breakthroughs with deep learning. I mentioned earlier that speech recognition came around 2010, and then computer vision around 2012. A couple of years later, in 2014 and 2015, we had breakthroughs in machine translation that ended up being used in Google Translate in 2016. 2016 was also the year we saw the breakthroughs with AlphaGo. All of these things, among a number of others, were really not expected.

我记得在 2014 年，我查看了我们在标题生成方面的一些成果，当时计算机正在尝试为图像想出一个标题，我很惊讶我们能够做到这一点。如果你在一年前问我我们是否能在一年内做到这一点，我会说不可能。

I remember back in 2014 I looked at some of our results in caption generation, where the computer is trying to come up with a caption for an image, and I was amazed that we were able to do that. If you had asked me just one year earlier if we’d be able to do that in a year, I would have said no.

马丁·福特：这些字幕非常精彩。有时它们离题万里，但大多数时候都令人惊叹。

MARTIN FORD: Those captions are pretty remarkable. Sometimes they’re way off the mark, but most of the time they’re amazing.

YOSHUA BENGIO：当然，它们有时也差得很远！它们没有接受足够的数据训练，而且这些系统还需要在基础研究方面取得一些根本性的进展，才能真正理解图像和语言。我们距离实现这些进步还很远，但它们能够达到目前的性能水平这一事实是我们意料之外的。

YOSHUA BENGIO: Of course, they’re way off sometimes! They’re not trained on enough data, and there are also some fundamental advances in basic research that need to be made for those systems to really understand an image and really understand language. We’re far away from achieving those advances, but the fact that they were able to reach the level of performance that they have was not something we expected.

马丁·福特：我们来谈谈你的职业生涯吧。你进入人工智能领域的历程是怎样的？

MARTIN FORD: Let’s talk about your career. What was your own path into the field of AI?

YOSHUA BENGIO：我年轻的时候读过很多科幻小说，我相信这对我产生了影响。它让我了解了人工智能和阿西莫夫的机器人三定律等话题，我想上大学学习物理和数学。当我和哥哥对电脑产生兴趣时，情况发生了变化。我们攒钱买了一台 Apple IIe，然后买了一台 Atari 800。当时软件很少，所以我们学会了用 BASIC 自己编程。

YOSHUA BENGIO: When I was young, I would read a lot of science fiction, and I’m sure that had an impact on me. It introduced me to topics such as AI and Asimov’s Three Laws of Robotics, and I wanted to go to college and study physics and mathematics. That changed when my brother and I became interested in computers. We saved our money to buy an Apple IIe and then an Atari 800. Software was scarce in those days, so we learned to program them ourselves in BASIC.

我对编程非常感兴趣，因此我攻读了计算机工程专业，然后攻读了计算机科学专业硕士和博士学位。在 1985 年左右攻读硕士学位期间，我开始阅读一些关于早期神经网络的论文，包括 Geoffrey Hinton 的一些论文，我对它一见钟情。我很快就决定这就是我想要研究的主题。

I got so excited with programming that I went into computer engineering and then computer science for my Master’s and PhD. While doing my Master’s around 1985, I started reading some papers on early neural nets, including some of Geoffrey Hinton’s papers, and it was like love at first sight. I quickly decided that this was the subject I wanted to do my research in.

马丁·福特：对于那些想要进入深度学习专家或研究人员领域的人，您有什么特别的建议吗？

MARTIN FORD: Is there any particular advice you’d give to someone who wants to get into the field of being a deep learning expert or researcher?

YOSHUA BENGIO：跳入水中开始游泳吧。由于人们对这个领域非常感兴趣，因此有大量信息，包括各种级别的教程、视频和开源库。我合著了一本书，名为《深度学习》，它可以帮助新手进入这个领域，并且可以在线免费获取。我看到许多本科生通过阅读大量论文、尝试重现这些论文来训练自己，然后申请进入进行此类研究的实验室。如果你对这个领域感兴趣，现在就是开始的最佳时机。

YOSHUA BENGIO: Just jump in the water and start swimming. There’s a ton of information in the form of tutorials, videos, and open source libraries at all levels because there’s so much interest in this field. And there is the book I co-authored, called Deep Learning, which helps newcomers into the field and is available for free online. I see many undergrad students training themselves by reading lots and lots of papers, trying to reproduce those papers, and then applying to get into the labs which are doing this kind of research. If you’re interested in the area, there’s no better time to start than now.

马丁·福特：就您的职业生涯而言，我注意到的一件事是，在深度学习领域的关键人物中，您是唯一一个完全留在学术界的人。其他人大多在 Facebook 或谷歌等公司兼职。是什么让您走上了这条职业道路？

MARTIN FORD: In terms of your career, one thing I noticed is that of the key people in deep learning, you’re the only one that remains entirely in the academic world. Most others are part-time at companies like Facebook or Google. What made you take that career pathway?

YOSHUA BENGIO：我一直很重视学术，以及为公共利益或我认为会产生更大影响的事情而工作的自由。我也重视与学生的合作，无论是从心理上还是从研究效率和生产力方面。如果我进入这个行业，我会抛弃很多这些。

YOSHUA BENGIO: I’ve always valued academia and the freedom to work for the common good or the things that I believe would have more impact. I also value working with students both psychologically and in terms of the efficiency and productivity of my research. If I went into the industry, I would be leaving a lot of that behind.

我也想留在蒙特利尔，当时的情况是，进入这个行业意味着要么去加利福尼亚，要么去纽约。正是在那时，我想也许我们可以在蒙特利尔建立一些东西，让它成为人工智能的新硅谷。因此，我决定留下来，创建蒙特利尔学习算法研究所 Mila。

I also wanted to stay in Montreal, and at that time, it was the case that going into the industry meant going to either California or New York. It was then that I thought that maybe we could build something in Montreal that could become a new Silicon Valley for AI. As a result, I decided to stay and create Mila, The Montreal Institute for Learning Algorithms.

Mila 负责基础研究，并在蒙特利尔的人工智能生态系统中发挥领导作用。该角色涉及与多伦多的 Vector Institute 和埃德蒙顿的 Amii 合作，作为加拿大真正推动人工智能发展的战略的一部分——在科学、经济和积极的社会影响方面。

Mila carries out basic research, and also plays a leadership role in the AI ecosystem in Montreal. This role involves working in partnership with the Vector Institute in Toronto, and Amii, in Edmonton, as part of the Canadian strategy to really push AI forward—in terms of science, in terms of the economy, and in terms of positive social impact.

马丁·福特：既然你提到了，那我们就多谈谈人工智能和经济，以及其中的一些风险。我写了很多关于人工智能可能引发新工业革命，并可能导致大量失业的文章。你对这个假设有什么看法？你认为它被夸大了吗？

MARTIN FORD: Since you mention it, let’s talk more about AI and the economy, and some of the risks there. I have written a lot about the potential for artificial intelligence to bring on a new Industrial Revolution, and potentially to lead to a lot of job losses. How do you feel about that hypothesis, do you think that it is overhyped?

YOSHUA BENGIO：不，我不认为这被夸大了。目前尚不清楚的是，这是否会在未来十年或三十年内实现。我可以说的是，即使我们明天停止人工智能和深度学习的基础研究，科学也已经足够先进，只需从这些想法中设计出新服务和新产品，就能从中收获巨大的社会和经济效益。

YOSHUA BENGIO: No, I don’t think it’s overhyped. The part that is less clear is whether this is going to happen over a decade or three decades. What I can say is that even if we stop basic research in AI and deep learning tomorrow, the science has advanced enough that there’s already a huge amount of social and economic benefit to reap from it simply by engineering new services and new products from these ideas.

我们还收集了大量我们不使用的数据。例如，在医疗保健领域，我们只使用了可用数据的极小部分，或者说，随着每天数字化的数据越来越多，我们可以使用的数据将越来越少。硬件公司正在努力打造深度学习芯片，这些芯片的速度很快将比我们目前使用的芯片快一千倍，能效也更高。这些东西可以在你身边的任何地方使用，无论是在汽车里还是在手机里，这一事实显然将改变世界。

We also collect a huge amount of data that we don’t use. For example, in healthcare, we’re only using a tiny, tiny fraction of what is available, or of what will be available as even more gets digitized every day. Hardware companies are working hard to build deep learning chips that are soon going to be easily a thousand times faster or more energy-efficient than the ones we currently have. The fact that you could have these things everywhere around you, in cars and phones, is clearly going to change the world.

社会因素等因素会拖慢进程。即使技术已经成熟，改变医疗基础设施也需要时间。即使技术在不断进步，社会也不可能无限快速地改变。

What will slow things down are things like social factors. It takes time to change the healthcare infrastructure, even if the technology is there. Society can’t change infinitely fast, even if the technology is moving forward.

马丁·福特：如果这项技术变革确实导致大量工作岗位消失，您认为基本收入之类的制度会是一个好的解决方案吗？

MARTIN FORD: If this technology change does lead to a lot of jobs being eliminated, do you think something like a basic income would be a good solution?

约书亚·本吉奥：我认为基本收入是可行的，但我们必须从科学的角度看待这个问题，摆脱我们的道德偏见，即如果一个人不工作，就不应该有收入。我认为这太疯狂了。我认为我们必须考虑什么对经济最有利，什么对人们的幸福最有利，我们可以进行试点实验来回答这些问题。

YOSHUA BENGIO: I think a basic income could work, but we have to take a scientific view on this to get rid of our moral priors that say if a person doesn’t work, then they shouldn’t have an income. I think it’s crazy. I think we have to look at what’s going to work best for the economy and what’s going to work best for people’s happiness, and we can do pilot experiments to answer those questions.

这并不是一个明确的答案，社会可以通过多种方式照顾那些被遗弃的人，并尽量减少工业革命带来的苦难。我要回顾我的朋友 Yann LeCun 说过的一句话：如果我们在 19 世纪就有远见，能看到工业革命将如何展开，也许我们就可以避免随后发生的许多苦难。如果我们在 19 世纪就建立了目前大多数西方国家都存在的那种社会安全网，而不是等到 20 世纪 40 年代和 50 年代，那么数亿人将过上更好、更健康的生活。问题是，这次可能要花不到一个世纪的时间才能展开这个故事，因此潜在的负面影响可能会更大。

It’s not like there’s one clear answer, there are many ways that society could take care of the people who are going to be left behind and minimize the amount of misery arising from this Industrial Revolution. I’m going to go back to something that my friend Yann LeCun said: If we had had the foresight in the 19th century to see how the Industrial Revolution would unfold, maybe we could have avoided much of the misery that followed. If in the 19th century we had put in place the kind of social safety net that currently exists in most Western nations, instead of waiting until the 1940s and 1950s, then hundreds of millions of people would have led a much better and healthier life. The thing is, it’s going to take probably much less than a century this time to unfold that story, and so the potential negative impacts could be even larger.

我认为现在就开始思考这个问题并开始科学地研究减少痛苦和改善全球福祉的方案非常重要。我认为这是可能的，我们不应该仅仅依靠我们过去的偏见和宗教信仰来决定这些问题的答案。

I think it’s really important to start thinking about it right now and to start scientifically studying the options to minimize misery and optimize global well-being. I think it’s possible to do it, and we shouldn’t just rely on our old biases and religious beliefs in order to decide on the answer to these questions.

马丁·福特：我同意，但正如你所说，情况可能会迅速发展。这也将是一个惊人的政治问题。

MARTIN FORD: I agree, but as you say, it could unfold fairly rapidly. It’s going to be a staggering political problem, too.

YOSHUA BENGIO：这更是我们要迅速采取行动的理由！

YOSHUA BENGIO: Which is all the more reason to act quickly!

马丁·福特：观点很有道理。除了经济影响之外，人工智能还有哪些值得我们担心的问题？

MARTIN FORD: A valid point. Beyond the economic impact, what are the other things we should worry about in terms of artificial intelligence?

约书亚·本吉奥：我一直积极反对杀手机器人。

YOSHUA BENGIO: I have been very active in speaking against killer robots.

马丁·福特：我注意到你签署了一封致韩国一所大学的信，该大学似乎正致力于杀手机器人的研究。

MARTIN FORD: I noticed you signed a letter aimed at a university in Korea which seemed to be headed towards research on killer robots.

YOSHUA BENGIO：没错，这封信确实起了作用。事实上，韩国科学技术研究院（KAIST）一直在告诉我们，他们将避免开发无人参与的军事系统。

YOSHUA BENGIO: That’s right, and this letter is working. In fact, KAIST, The Korea Advanced Institute of Science and Technology, has been telling us that they will avoid going into the development of military systems which don’t have a human in the loop.

让我回到关于人类参与的问题，因为我认为这真的很重要。人们需要明白，当前的人工智能——以及我们可以预见的未来的人工智能——没有、也不会有道德感或对什么是对什么是错的道德理解。我知道不同文化之间存在差异，但这些道德问题对人们的生活很重要。

Let me go back to this question about a human in the loop because I think this is really important. People need to understand that current AI—and the AI that we can foresee in the reasonable future—does not, and will not, have a moral sense or moral understanding of what is right and what is wrong. I know there are differences across cultures, but these moral questions are important in people’s lives.

确实如此，不只是对于杀手机器人而言，对于其他所有事物而言，比如法官决定一个人的命运的工作——这个人是否应该回到监狱还是被释放到社会。这些都是非常困难的道德问题，你必须了解人类的心理，你必须了解道德价值观。将这些决定权交给没有这种理解能力的机器是疯狂的。这不仅疯狂，而且是错误的。我们必须制定社会规范或法律，确保在可预见的未来计算机不会承担这样的责任。

It’s true, not just for killer robots but all kinds of other things, like the work that a judge does deciding on the fate of a person—whether that person should return to prison or be freed into society. These are really difficult moral questions, where you have to understand human psychology, and you have to understand moral values. It’s crazy to put those decisions in the hands of machines, which don’t have that kind of understanding. It’s not just crazy; it’s wrong. We have to have social norms or laws, which make sure that computers in the foreseeable future don’t get those kinds of responsibilities.

马丁·福特：我想就此向您提出挑战。我想很多人会说您对人类及其判断力的质量有着非常理想化的看法。

MARTIN FORD: I want to challenge you on that. I think a lot of people would say that you have a very idealistic view of human beings and the quality of their judgment.

约书亚·本吉奥：当然，但我宁愿让不完美的人来担任法官，也不愿让机器不明白自己在做什么。

YOSHUA BENGIO: Sure, but I’d rather have an imperfect human being as a judge than a machine that doesn’t understand what it’s doing.

马丁·福特：但想象一下，一个自主安全机器人会很乐意先挡子弹，然后再开枪，而人类永远不会这样做，而且这可能会挽救生命。理论上，如果编程正确，自主安全机器人也不会有种族歧视。这些实际上是它可能比人类更有优势的领域。你同意吗？

MARTIN FORD: But think of an autonomous security robot that would be happy to take a bullet first and shoot second, whereas a human would never do that, and that could potentially save lives. In theory, an autonomous security robot would also not be racist, if it were programmed correctly. These are actually areas where it might have an advantage over a human being. Would you agree?

YOSHUA BENGIO：也许有一天会这样，但我可以告诉你，我们还没有达到那个程度。这不仅关乎精确度，还关乎理解人类的背景，而计算机对此一无所知。

YOSHUA BENGIO: Well, it might be the case one day, but I can tell you we’re not there yet. It’s not just about precision, it’s about understanding the human context, and computers have absolutely zero clues about that.

马丁·福特：除了军事和武器化方面，我们还应该担心人工智能的其他什么问题吗？

MARTIN FORD: Other than the military and weaponization aspects, is there anything else that we should be worried about with AI?

YOSHUA BENGIO：是的，这个问题之前没有被讨论过太多，但现在可能会因为 Facebook 和剑桥分析公司事件而变得更加突出。我们应该真正意识到，在广告中使用人工智能或一般性地影响人们的行为对民主来说是危险的，而且在某些方面在道德上是错误的。我们应该确保我们的社会尽可能地防止这些事情发生。

YOSHUA BENGIO: Yes, and this is something that hasn’t been discussed much, but now may come more to the forefront because of what happened with Facebook and Cambridge Analytica. The use of AI in advertising or generally in influencing people is something that we should be really aware of as dangerous for democracy—and is morally wrong in some ways. We should make sure that our society prevents those things as much as possible.

例如，在加拿大，针对儿童的广告是被禁止的。这样做有一个很好的理由：我们认为，在儿童如此脆弱的情况下操纵他们的思想是不道德的。但事实上，我们每个人都是脆弱的，如果不是这样，那么广告就不起作用了。

In Canada, for example, advertising that is directed at children is forbidden. There’s a good reason for that: We think that it’s immoral to manipulate their minds when they are so vulnerable. In fact, though, every one of us is vulnerable, and if it weren’t the case, then advertising wouldn’t work.

另一件事是，广告实际上会损害市场力量，因为它为大公司提供了一种工具，可以减缓小公司进入其市场的速度，因为大公司可以使用他们的品牌。如今，他们可以使用人工智能以更准确的方式向人们传达信息，我认为这有点可怕，尤其是当它让人们做出可能不利于他们福祉的事情时。例如，政治广告可能就是这种情况，或者广告可能会改变你的行为并影响你的健康。我认为我们应该非常非常小心地使用这些工具来影响普通人。

The other thing is that advertising actually hurts market forces because it gives larger companies a tool to slow down smaller companies coming into their markets because those larger companies can use their brand. Nowadays they can use AI to target their message to people in a much more accurate way, and I think that’s kind of scary, especially when it makes people do things that may be against their well-being. It could be the case in political advertising, for example, or advertising that could change your behavior and have an impact on your health. I think we should be really, really careful about how these tools are used to influence people in general.

马丁·福特：伊隆·马斯克和斯蒂芬·霍金等人警告说，超级人工智能将对人类造成生存威胁，并陷入不断循环的恶性循环，对此您怎么看？这些是我们现在应该担心的事情吗？

MARTIN FORD: What about the warnings from people like Elon Musk and Stephen Hawking about an existential threat from super intelligent AI and getting into a recursive improvement loop? Are these things that we should be concerned about at this point?

YOSHUA BENGIO：我并不担心这些事情，我认为有人研究这个问题是可以的。我对当前科学的理解以及我可以预见的结果是，这些情况是不现实的。这些情况与我们目前构建人工智能的方式不相容。几十年后情况可能会有所不同，我不知道，但就我而言，这只是科幻小说。我认为也许这些担忧会分散我们现在可以采取行动解决的一些最紧迫问题的注意力。

YOSHUA BENGIO: I’m not concerned about these things, I think it’s fine that some people study the question. My understanding of the current science as it is now, and as I can foresee it, is that those kinds of scenarios are not realistic. Those kinds of scenarios are not compatible with how we build AI right now. Things may be different in a few decades, I have no idea, but that is science fiction as far as I’m concerned. I think perhaps those fears are detracting from some of the most pressing issues that we could act on now.

我们讨论过杀手机器人，也讨论过政治广告，但还有其他问题，比如数据可能存在偏见并加剧歧视。这些都是政府和公司现在可以采取行动的问题，我们确实有一些方法来缓解其中一些问题。辩论不应该过多地关注这些非常长期的潜在风险，我认为这与我对人工智能的理解不相符，但我们应该关注杀手机器人等短期问题。

We’ve talked about killer robots and we’ve talked about political advertising, but there are other concerns, like how data could be biased and reinforce discrimination, for example. These are things that governments and companies can act on now, and we do have some ways to mitigate some of these issues. The debate shouldn’t focus so much on these very long-term potential risks, which I don’t think are compatible with my understanding of AI, but we should pay attention to short-term things like killer robots.

马丁·福特：我想问您关于与中国和其他国家的潜在竞争的问题。例如，您谈了很多关于对自主武器的限制，其中一个明显的担忧是，一些国家可能会忽视这些规则。我们应该对这种国际竞争有多担心？

MARTIN FORD: I want to ask you about the potential competition with China and other countries. You’ve talked a lot about, for example, having limitations on autonomous weapons and one obvious concern there is that some countries might ignore those rules. How worried should we be about that international competition?

YOSHUA BENGIO：首先，在科学方面我并不担心。全世界从事科学研究的研究人员越多，对科学就越有利。如果中国在人工智能方面投入大量资金，那很好；归根结底，我们都会利用这项研究带来的进步。

YOSHUA BENGIO: Firstly, on the scientific side I don’t have any concern. The more researchers around the world are working on a science, the better it is for that science. If China is investing a lot in AI that’s fine; at the end of the day, we’re all going to take advantage of the progress that’s going to come of that research.

然而，我认为中国政府可能将这项技术用于军事目的或内部警务，这一点令人担忧。如果你利用当前的科学水平，建立能够识别人、识别面孔和跟踪他们的系统，那么基本上你可以在短短几年内建立一个“老大哥”社会。这在技术上是相当可行的，而且它正在给世界各地的民主带来更大的危险。这确实是一件值得担忧的事情。不仅像中国这样的国家可能发生这种情况；如果自由民主国家滑向独裁统治，这种情况也可能会发生，正如我们在一些国家看到的那样。

However, I think the part about the Chinese government potentially using this technology either for military purposes or for internal policing is scary. If you take the current state of the science and build systems that will recognize people, recognize faces, and track them, then essentially you can build a Big Brother society in just a few years. It’s quite technically feasible and it is creating even more danger for democracy around the world. That is really something to be concerned about. It’s not just states like China where this could happen, either; it could also happen in liberal democracies if they slip towards autocratic rule, as we have seen in some countries.

关于军事上人工智能的竞赛，我们不应该将杀手机器人与军事上人工智能的使用混为一谈。我并不是说我们应该完全禁止军事上人工智能的使用。例如，如果军方使用人工智能制造能够摧毁杀手机器人的武器，那么这是一件好事。不道德的是让这些机器人杀死人类。我们并不是都必须不道德地使用人工智能。我们可以制造防御性武器，这可能有助于阻止这场竞赛。

Regarding the military race to use AI, we shouldn’t confuse killer robots with the use of AI in the military. I’m not saying that we should completely ban the use of AI in the military. For example, if the military uses AI to build weapons that will destroy killer robots, then that’s a good thing. What is immoral is to have these robots kill humans. It’s not like we all have to use AI immorally. We can build defensive weapons, and that could be useful to stop the race.

马丁·福特：听起来你觉得在自主武器方面监管肯定发挥着作用？

MARTIN FORD: It sounds like you feel there’s definitely a role for regulation in terms of autonomous weapons?

YOSHUA BENGIO：监管无处不在。在人工智能将产生社会影响的领域，我们至少必须考虑监管。我们必须考虑什么是正确的社会机制，以确保人工智能得到良好的利用。

YOSHUA BENGIO: There’s a role for regulation everywhere. In the areas where AI is going to have a social impact, then we at least have to think about regulation. We have to consider what the right social mechanism is that will make sure that AI is used for good.

马丁·福特：您认为政府有能力回答这个问题吗？

MARTIN FORD: And you think governments are equipped to take on that question?

YOSHUA BENGIO：我不相信公司会自己做到这一点，因为他们的主要关注点是实现利润最大化。当然，他们也试图在用户或客户中保持受欢迎程度，但他们对自己所做的事情并不完全透明。他们所实现的目标是否符合整个民众的福祉并不总是很清楚。

YOSHUA BENGIO: I don’t trust companies to do it by themselves because their main focus is on maximizing profits. Of course, they’re also trying to remain popular among their users or customers, but they’re not completely transparent about what they do. It’s not always clear that those objectives that they’re implementing correspond to the well-being of the population in general.

我认为政府发挥着非常重要的作用，而且不只是个别政府，而是国际社会，因为其中许多问题不仅仅是地方问题，而是国际问题。

I think governments have a really important role to play, and it’s not just individual governments, it’s the international community because many of these questions are not just local questions, they’re international questions.

马丁·福特：您是否认为这一切带来的好处将明显大于风险？

MARTIN FORD: Do you believe that the benefits to all of this are going to clearly outweigh the risks?

约书亚·本吉奥：只有我们明智行事，风险才能大于风险。这就是为什么进行这些讨论如此重要。这就是为什么我们不想带着眼罩直行；我们必须睁大眼睛，警惕所有潜在的潜在危险。

YOSHUA BENGIO: They’ll only outweigh the risks if we act wisely. That’s why it’s so important to have those discussions. That’s why we don’t want to move straight ahead with blinkers on; we have to keep our eyes open to all of the potential dangers that are lurking.

马丁·福特：您认为现在应该在哪里进行这一讨论？这主要是智库和大学应该做的事情吗？还是您认为这应该成为国内和国际政治讨论的一部分？

MARTIN FORD: Where do you think this discussion should be taking place now? Is it something primarily think tanks and universities should do, or do you think this should be part of the political discussion both nationally and internationally?

YOSHUA BENGIO：这完全应该成为政治讨论的一部分。我受邀在七国集团部长会议上发言，讨论的问题之一是“我们如何以既有利于经济又能保持人民信任的方式发展人工智能？”因为今天的人们确实有顾虑。答案是不要秘密或在象牙塔里做事，而是进行公开讨论，让桌子周围的每个人，包括每个公民，都参与讨论。我们将不得不集体选择我们想要什么样的未来，而且由于人工智能如此强大，每个公民都应该在某种程度上了解问题是什么。

YOSHUA BENGIO: It should totally be part of the political discussion. I was invited to speak at a meeting of G7 ministers, and one of the questions discussed was, “How do we develop AI in a way that’s both economically positive and keeps the trust of the people?”, because people today do have concerns. The answer is to not do things in secret or in ivory towers, but instead to have an open discussion where everybody around the table, including every citizen, should be part of the discussion. We’re going to have to make collective choices about what kind of future we want, and because AI is so powerful, every citizen should understand at some level what the issues are.

YOSHUA BENGIO 是计算机科学与运筹学系的全职教授、蒙特利尔学习算法研究所 (Mila) 的科学主任、CIFAR 机器和大脑学习计划的联合主任、加拿大统计学习算法研究主席。他与 Ian Goodfellow 和 Aaron Courville 共同编写了《深度学习》，这是该主题的代表教科书之一。该书可从https://www.deeplearningbook.org免费获取。

YOSHUA BENGIO is Full Professor of the Department of Computer Science and Operations Research, scientific director of the Montreal Institute for Learning Algorithms (Mila), CIFAR Program co-director of the CIFAR program on Learning in Machines and Brains, Canada Research Chair in Statistical Learning Algorithms. Together with Ian Goodfellow and Aaron Courville, he wrote Deep Learning, one of the defining textbooks on the subject. The book is available for free from https://www.deeplearningbook.org.

加州大学伯克利分校计算机科学教授

PROFESSOR OF COMPUTER SCIENCE, UNIVERSITY OF CALIFORNIA, BERKELEY

Stuart J. Russell 被公认为人工智能领域的世界领先贡献者之一。他是加州大学伯克利分校计算机科学教授兼人类兼容人工智能中心主任。Stuart 是领先的人工智能教科书《人工智能：一种现代方法》的合著者，该书在全球 1,300 多所学院和大学中使用。

Stuart J. Russell is widely recognized as one of the world’s leading contributors in the field of artificial intelligence. He is a Professor of Computer Science and Director of the Center for Human-Compatible Artificial Intelligence at The University of California, Berkeley. Stuart is the co-author of the leading AI textbook, Artificial Intelligence: A Modern Approach, which is in use at over 1,300 colleges and universities throughout the world.

马丁·福特：鉴于您是当今使用的标准人工智能教科书的合著者，我认为如果您能定义一些关键的人工智能术语，可能会很有趣。您对人工智能的定义是什么？它包括什么？该领域包括哪些类型的计算机科学问题？您能将它与机器学习进行比较或对比吗？

MARTIN FORD: Given that you co-wrote the standard textbook on AI in use today, I thought it might be interesting if you could define some key AI terms. What is your definition of artificial intelligence? What does it encompass? What types of computer science problems would be included in that arena? Could you compare it or contrast it with machine learning?

斯图尔特·J·拉塞尔：我来给你一个人工智能的标准定义，这个定义和书中的很相似，现在已被广泛接受：一个实体的智能程度取决于它做正确的事情，也就是说，它的行为有望实现其目标。这个定义适用于人类和机器。做正确的事情这一概念是人工智能的关键统一原则。当我们分解这一原则并深入研究在现实世界中做正确的事情需要什么时，我们意识到一个成功的人工智能系统需要一些关键能力，包括感知、视觉、语音识别和行动。

STUART J. RUSSELL: Let me give you, shall we say, the standard definition of artificial intelligence, which is similar to the one in the book and is now quite widely accepted: An entity is intelligent to the extent that it does the right thing, meaning that its actions are expected to achieve its objectives. The definition applies to both humans and machines. This notion of doing the right thing is the key unifying principle of AI. When we break this principle down and look deeply at what is required to do the right thing in the real world, we realize that a successful AI system needs some key abilities, including perception, vision, speech recognition, and action.

这些能力有助于我们定义人工智能。我们谈论的是控制机器人操纵器的能力，以及机器人技术中发生的一切。我们谈论的是决策、规划和解决问题的能力。我们谈论的是沟通的能力，因此自然语言理解对人工智能也变得极为重要。

These abilities help us to define artificial intelligence. We’re talking about the ability to control robot manipulators, and everything that happens in robotics. We’re talking about the ability to make decisions, to plan, and to problem-solve. We’re talking about the ability to communicate, and so natural language understanding also becomes extremely important to AI.

我们还谈论了内在了解事物的能力。如果你实际上什么都不知道，那么在现实世界中很难成功运作。为了理解我们如何了解事物，我们进入了我们称之为知识表示的科学领域。在这里，我们研究知识如何被内部存储，然后通过推理算法（例如自动逻辑推理和概率推理算法）进行处理。

We’re also talking about an ability to internally know things. It’s very hard to function successfully in the real world if you don’t actually know anything. To understand how we know things, we enter the scientific field that we call knowledge representation. This is where we study how knowledge can be stored internally and then processed by reasoning algorithms, such as automated logical deduction and probabilistic inference algorithms.

然后是学习。学习是现代人工智能的一项关键能力。机器学习一直是人工智能的一个分支，它只是意味着提高你根据经验做正确事情的能力。这可能是学习如何通过观察带标签的物体示例来更好地感知。这也可能意味着学习如何通过经验更好地推理——例如发现哪些推理步骤对解决问题有用，哪些推理步骤用处不大。

Then there is learning. Learning is a key ability for modern artificial intelligence. Machine learning has always been a subfield of AI, and it simply means improving your ability to do the right thing as a result of experience. That could be learning how to perceive better by seeing labeled examples of objects. That could also mean learning how to reason better by experience—such as discovering which reasoning steps turn out to be useful for solving a problem, and which reasoning steps turn out to be less useful.

例如，AlphaGo 是一款现代人工智能围棋程序，最近击败了人类世界冠军，它确实在学习。它学会了如何从经验中更好地推理。除了学习评估位置之外，AlphaGo 还学习如何控制自己的思考，以便能够更有效地更快地做出高质量的决策，同时减少计算量。

AlphaGo, for example, is a modern AI Go program that recently beat the best human world-champion players, and it really does learn. It learns how to reason better from experience. As well as learning to evaluate positions, AlphaGo learns how to control its own deliberations so that it more effectively reaches high decision-quality moves more quickly, with less computation.

马丁·福特：您能定义一下神经网络和深度学习吗？

MARTIN FORD: Can you also define neural networks and deep learning?

斯图尔特·J·拉塞尔：是的，在机器学习中，一种标准技术被称为“监督学习”，即我们为人工智能系统提供一组概念示例，以及该集合中每个示例的描述和标签。例如，我们可能有一张照片，其中有图像中的所有像素，然后我们有一个标签，说明这是一张船的照片，或一只达尔马提亚狗的照片，或一碗樱桃的照片。在这项任务的监督学习中，目标是找到一个预测因子或假设，以对图像进行一般分类。

STUART J. RUSSELL: Yes, in machine learning one of the standard techniques is called “supervised learning,” where we give the AI system a set of examples of a concept, along with a description and a label for each example in the set. For example, we might have a photograph, where we’ve got all the pixels in the image, and then we have a label saying that this is a photograph of a boat, or of a Dalmatian dog, or of a bowl of cherries. In supervised learning for this task, the goal is to find a predictor, or a hypothesis, for how to classify images in general.

通过这些监督训练示例，我们尝试让人工智能具备识别达尔马提亚犬图片的能力，以及预测其他达尔马提亚犬图片的外观的能力。

From these supervised training examples, we try to give an AI the ability to recognize pictures of, say, Dalmatian dogs, and the ability to predict how other pictures of Dalmatian dogs might look.

表示假设或预测因子的一种方式是神经网络。神经网络本质上是一个具有多层的复杂电路。该电路的输入可以是达尔马提亚犬图片中的像素值。然后，随着这些输入值在电路中传播，电路的每一层都会计算出新的值。最后，我们得到了神经网络的输出，即对正在识别哪种物体的预测。

One way of representing the hypothesis, or the predictor, is a neural net. A neural net is essentially a complicated circuit with many layers. The input into this circuit could be the values of pixels from pictures of Dalmatian dogs. Then, as those input values propagate through the circuit, new values are calculated at each layer of the circuit. At the end, we have the outputs of the neural network, which are the predictions about what kind of object is being recognized.

因此，如果我们的输入图像中有一只斑点狗，那么当所有这些数字和像素值通过神经网络及其所有层和连接时，斑点狗的输出指示器将亮起高值，而一碗樱桃的输出指示器将亮起低值。然后我们说神经网络已正确识别出斑点狗。

So hopefully, if there’s a Dalmatian dog in our input image, then by the time all those numbers and pixel values propagate through the neural network and all of its layers and connections, the output indicator for a Dalmatian dog will light up with a high value, and the output indicator for a bowl of cherries will have a low value. We then say that the neural network has correctly recognized a Dalmatian dog.

马丁·福特：如何让神经网络识别图像？

MARTIN FORD: How do you get a neural network to recognize images?

斯图尔特·J·拉塞尔：这就是学习过程的开始。电路中的所有连接之间都有可调节的连接强度，而学习算法所做的就是调整这些连接强度，以便网络倾向于对训练示例做出正确的预测。然后，如果你幸运的话，神经网络也会对它以前从未见过的新图像做出正确的预测。这就是神经网络！

STUART J. RUSSELL: This is where the learning process comes in. The circuit has adjustable connection strengths between all its connections, and what the learning algorithms do is adjust those connection strengths so that the network tends to give the correct predictions on the training examples. Then if you’re lucky, the neural network will also give correct predictions on new images that it hasn’t seen before. And that’s a neural network!

更进一步来说，深度学习就是拥有多层神经网络。神经网络的深度没有最低要求，但我们通常认为两层或三层不是深度学习网络，而四层或以上才是深度学习。

Going one step further, deep learning is where we have neural networks that have many layers. There is no required minimum for a neural network to be deep, but we would usually say that two or three layers is not a deep learning network, while four or more layers is deep learning.

一些深度学习网络多达一千层或更多。通过在深度学习中设置多层，我们可以表示输入和输出之间非常复杂的转换，通过组合更简单的转换，每个转换由网络中的一层表示。

Some deep learning networks get up to one thousand layers or more. By having many layers in deep learning, we can represent a very complex transformation between the input and output, by a composition of much simpler transformations, each represented by one of those layers in the network.

深度学习假设认为，多层结构使得学习算法更容易找到预测器，并设置网络中的所有连接强度，从而使其更好地完成工作。

The deep learning hypothesis suggests that many layers make it easier for the learning algorithm to find a predictor, to set all the connection strengths in the network so that it does a good job.

我们现在才刚刚开始从理论上理解深度学习假设何时以及为何正确，但在很大程度上，它仍然是一种魔法，因为它实际上并不一定非得这样发生。现实世界中的图像似乎具有某种属性，现实世界中的声音和语音信号也具有某种属性，因此当你将这类数据连接到深度网络时，出于某种原因，学习一个好的预测器会相对容易。但为什么会发生这种情况，人们仍不得而知。

We are just beginning now to get some theoretical understanding of when and why the deep learning hypothesis is correct, but to a large extent, it’s still a kind of magic, because it really didn’t have to happen that way. There seems to be a property of images in the real world, and there is some property of sound and speech signals in the real world, such that when you connect that kind of data to a deep network it will—for some reason—be relatively easy to learn a good predictor. But why this happens is still anyone’s guess.

马丁·福特：深度学习现在受到广泛关注，人们很容易认为人工智能就是深度学习的同义词。但深度学习实际上只是该领域相对较小的一部分，不是吗？

MARTIN FORD: Deep learning is receiving enormous amounts of attention right now, and it would be easy to come away with the impression that artificial intelligence is synonymous with deep learning. But deep learning is really just one relatively small part of the field, isn’t it?

斯图尔特·J·拉塞尔：是的，如果有人认为深度学习与人工智能是一回事，那就大错特错了，因为区分达尔马提亚犬和樱桃碗的能力很有用，但这只是我们为使人工智能取得成功而需要赋予它的很小一部分。感知和图像识别都是在现实世界中成功运作的重要方面，但深度学习只是其中的一部分。

STUART J. RUSSELL: Yes, it would be a huge mistake for someone to think that deep learning is the same thing as artificial intelligence, because the ability to distinguish Dalmatian dogs from bowls of cherries is useful but it is still only a very small part of what we need to give an artificial intelligence in order for it to be successful. Perception and image recognition are both important aspects of operating successfully in the real world, but deep learning is only one part of the picture.

AlphaGo 及其继任者 AlphaZero 在围棋和国际象棋方面取得了惊人的进步，引起了媒体对深度学习的广泛关注，但它们实际上是传统基于搜索的人工智能与深度学习算法的混合体，深度学习算法会评估传统人工智能系统搜索到的每个游戏位置。虽然区分好位置和坏位置的能力是 AlphaGo 的核心，但它无法仅通过深度学习就达到世界冠军级别的围棋水平。

AlphaGo, and its successor AlphaZero, created a lot of media attention around deep learning with stunning advances in Go and Chess, but they’re really a hybrid of classical search-based AI and a deep learning algorithm that evaluates each game position that the classical AI system searches through. While the ability to distinguish between good and bad positions is central to AlphaGo, it cannot play world-champion-level Go just by deep learning.

自动驾驶汽车系统还采用了传统基于搜索的人工智能和深度学习的混合体。自动驾驶汽车不仅仅是纯粹的深度学习系统，因为这种系统效果并不好。许多驾驶情况需要经典规则才能使人工智能取得成功。例如，如果你在中间车道，想要向右变道，而有人试图从内侧超车，那么你应该等他们先过去再靠边停车。对于需要预判的道路情况，由于没有令人满意的规则，可能需要想象汽车可能采取的各种行动以及其他汽车可能采取的各种行动，然后决定这些结果是好是坏。

Self-driving car systems also use a hybrid of classical search-based AI and deep learning. Self-driving cars are not just pure deep learning systems, because that does not work very well. Many driving situations need classical rules for an AI to be successful. For example, if you’re in the middle lane and you want to change lanes to the right, and there’s someone trying to pass you on the inside, then you should wait for them to go by first before you pull over. For road situations that require lookahead, because no satisfactory rule is available, it may be necessary to imagine various actions that the car could take as well as the various actions that other cars might take, and then decide if those outcomes are good or bad.

虽然感知非常重要，深度学习非常适合感知，但我们需要赋予人工智能系统多种不同类型的能力。当我们谈论跨越长期的活动时尤其如此，比如度假。或者建造工厂等非常复杂的行动。这些活动不可能由纯粹的深度学习黑箱系统来协调。

While perception is very important, and deep learning lends itself well to perception, there are many different types of ability that we need to give an AI system. This is particularly true when we’re talking about activities that span over long timescales, like going on a vacation. Or very complex actions like building a factory. There’s no possibility that those kinds of activities can be orchestrated by purely deep learning black-box systems.

让我以工厂为例来结束我关于深度学习局限性的观点。假设我们尝试使用深度学习来建造一家工厂。（毕竟，我们人类知道如何建造工厂，不是吗？）因此，我们将利用数十亿个以前的工厂建造示例来训练深度学习算法；我们将向它展示人类建造工厂的所有方式。我们将所有数据放入深度学习系统中，然后它就知道如何建造工厂了。我们能做到吗？不，这只是一个完全的白日梦。没有这样的数据，即使我们有数据，试图用这种方式建造工厂也没有任何意义。

Let me take the factory example to close my point about the limitations of deep learning here. Let’s imagine we try to use deep learning to build a factory. (After all, we humans know how to build a factory, don’t we?) So, we’ll take billions of previous examples of building factories to train a deep learning algorithm; we’ll show it all the ways that people have built factories. We take all that data and we put it into a deep learning system and then it knows how to build factories. Could we do that? No, it’s just a complete pipe dream. There is no such data, and it wouldn’t make any sense, even if we had it, to try to build factories that way.

我们需要知识来建造工厂。我们需要能够制定计划。我们需要能够推理物理障碍物和建筑物的结构特性。我们可以构建人工智能系统来解决这些现实世界的问题，但这不是通过深度学习来实现的。建造工厂需要完全不同类型的人工智能。

We need knowledge to build factories. We need to be able to construct plans. We need to be able to reason about physical obstructions and the structural properties of the buildings. We can build AI systems to work out these real-world problems, but it isn’t achieved by deep learning. Building a factory requires a different type of AI altogether.

马丁·福特：人工智能领域最近有哪些进展让您觉得不仅仅是渐进式的？您认为目前该领域处于绝对前沿的是什么？

MARTIN FORD: Are there recent advances in AI that have struck you as being more than just incremental? What would you point to that is at the absolute forefront of the field right now?

斯图尔特·J·拉塞尔：这个问题问得很好，因为目前新闻中的很多事情都不是真正的概念突破，它们只是演示。深蓝战胜卡斯帕罗夫就是一个完美的例子。深蓝基本上是 30 年前设计的算法的演示，这些算法逐渐得到增强，然后部署在越来越强大的硬件上，直到它们能够击败世界象棋冠军。但深蓝背后真正的概念突破在于如何设计象棋程序：前瞻如何工作；用于减少必须进行的搜索量的 alpha-beta 算法；以及设计评估函数的一些技术。因此，媒体经常将深蓝战胜卡斯帕罗夫描述为一项突破，而事实上，这一突破早在几十年前就发生了。

STUART J. RUSSELL: It’s a good question, because a lot of the things that are in the news at the moment are not really conceptual breakthroughs, they are just demos. The chess victory of Deep Blue over Kasparov is a perfect example. Deep Blue was basically a demo of algorithms that were designed 30 years earlier and had been gradually enhanced and then deployed on increasingly powerful hardware, until they could beat a world chess champion. But the actual conceptual breakthroughs behind Deep Blue were in how to design a chess program: how the lookahead works; the alpha-beta algorithm for reducing the amount of searching that had to be done; and some of the techniques for designing the evaluation functions. So, as is often the case, the media described the victory of Deep Blue over Kasparov as a breakthrough when in fact, the breakthrough had occurred decades earlier.

同样的事情今天仍在发生。例如，最近很多关于感知和语音识别的人工智能报告，以及关于听写准确度接近或超过人类听写准确度的头条新闻，都是非常令人印象深刻的实际工程成果，但它们再次展示了更早出现的概念突破——从可以追溯到 80 年代末和 90 年代初的早期深度学习系统和卷积网络。

The same thing is still happening today as well. For instance, a lot of the recent AI reports about perception and speech recognition, and headlines about dictation accuracy being close to or exceeding human dictation accuracy, are all very impressive practical engineering results, but they are again demos of conceptual breakthroughs that happened much earlier—from the early deep learning systems and convolutional networks that date right back to the late ‘80s and early ‘90s.

令人惊讶的是，几十年前我们就拥有了成功实现感知的工具；只是我们没有正确使用它们。通过将现代工程技术应用于较早的突破，通过收集大量数据集并在最新硬件上的超大网络上处理它们，我们最近成功地引起了人们对人工智能的极大兴趣，但这些并不一定处于人工智能的真正前沿。

It’s been something of a surprise that we already had the tools decades ago to do perception successfully; we just weren’t using them properly. By applying modern engineering to older breakthroughs, by collecting large datasets and processing them across very large networks on the latest hardware, we’ve managed to create a lot of interest recently in AI, but these have not necessarily been at the real forefront of AI.

马丁·福特：您认为 DeepMind 的 AlphaZero 是人工智能研究前沿技术的典范吗？

MARTIN FORD: Do you think DeepMind’s AlphaZero is a good example of a technology that’s right on the frontier of AI research?

斯图尔特·J·拉塞尔：我认为 AlphaZero 很有趣。对我来说，使用与下围棋相同的基本软件也能在世界冠军级别下国际象棋和将棋，这并不特别令人惊讶。所以，从这个意义上说，它并不是人工智能的前沿。

STUART J. RUSSELL: I think AlphaZero was interesting. To me, it was not particularly a surprise that you could use the same basic software that played Go to also play chess and Shogi at world-champion level. So, it was not at the forefront of AI in that sense.

我的意思是，当你想到 AlphaZero 在不到 24 小时的时间内，使用同一款软件在三款不同的游戏中学会了超人水平的水平时，你肯定会停下来思考。但这更证明了一种人工智能方法的正确性，即如果你对问题类型有清晰的理解，尤其是确定性、双人、轮流、完全可观察且规则已知的游戏，那么这类问题就可以用一类设计良好的人工智能算法来解决。这些算法已经存在了一段时间——可以学习好的评估函数并使用经典方法来控制搜索的算法。

I mean, it certainly gives you pause when you think that AlphaZero, in the space of less than twenty-four hours, learned to play at superhuman levels in three different games using the same software. But that’s more a vindication of an approach to AI that says that if you have a clear understanding of the problem class, especially deterministic, two-player, turn-taking, fully-observable games with known rules, then those kinds of problems are amenable to a well-designed class of AI algorithms. And these algorithms have been around for some time—algorithms that can learn good evaluation functions and use classical methods for controlling search.

显然，如果你想将这些技术扩展到其他类型的问题，你就必须想出不同的算法结构。例如，部分可观测性（即你看不到棋盘）需要不同类型的算法。例如，AlphaZero 无法玩扑克或驾驶汽车。这些任务需要一个能够估计看不见的东西的人工智能系统。AlphaZero 假设棋盘上的棋子就是棋盘上的棋子，仅此而已。

It’s also clear that if you want to extend those techniques to other classes of problems, you’re going to have to come up with different algorithmic structures. For example, partial observability—meaning that you can’t see the board, so to speak—requires a different class of algorithm. There’s nothing AlphaZero can do to play poker, for example, or to drive a car. Those tasks require an AI system that can estimate things that it can’t see. AlphaZero assumes that the pieces on the board are the pieces on the board, and that’s that.

马丁·福特：卡内基梅隆大学还开发了一款扑克牌 AI 系统，名为 Libratus？他们在那里实现了真正的 AI 突破吗？

MARTIN FORD: There was also a poker playing AI system developed at Carnegie Mellon University, called Libratus? Did they achieve a genuine AI breakthrough there?

斯图尔特·J·拉塞尔：卡内基梅隆大学的 Libratus 扑克 AI 是另一个非常令人印象深刻的混合 AI 示例：它是过去 10 到 15 年的研究成果中拼凑起来的几种不同算法的结合。在处理像扑克这样的部分信息游戏方面已经取得了很大进展。在像扑克这样的部分信息游戏中，会发生的情况之一是，你必须有一个随机的游戏策略，因为如果你总是虚张声势，那么人们就会发现你在虚张声势，然后他们就会揭穿你的虚张声势。但如果你从不虚张声势，那么当你手牌很弱时，你就永远无法从对手手中偷走一局。因此，人们早就知道，对于这类纸牌游戏，你应该随机化你的游戏行为，并以一定的概率虚张声势。

STUART J. RUSSELL: Carnegie Mellon’s Libratus poker AI was another very impressive hybrid AI example: it was a combination of several different algorithmic contributions that were pieced together from research that’s happened over the last 10 or 15 years. There has been a lot of progress in dealing with games like poker, which are games of partial information. One of the things that happens with partial-information games, like poker, is that you must have a randomized playing strategy because if, say, you always bluff, then people figure out that you’re bluffing and then they call your bluff. But if you never bluff, then you can never steal a game from your opponent when you have a weak hand. It’s long been known, therefore, that for these kinds of card games, you should randomize your playing behavior, and bluff with a certain probability.

玩好扑克的关键是调整下注的概率；也就是说，多久下注超过手牌实际价值的金额，多久下注较少的金额。这些概率的计算对于人工智能来说是可行的，而且可以非常精确地完成，但仅限于小型扑克版本，例如一副牌中只有几张牌。人工智能很难准确地对整场扑克游戏进行这些计算。因此，在人们致力于扩大扑克规模的十多年里，我们逐渐看到了计算这些概率的准确性和效率的提高，这些概率适用于越来越大的扑克版本。

The key to playing poker extremely well is adjusting those probabilities for how to bet; that is, how often to bet more than your hand really justifies, and how often to bet less. The calculations for these probabilities are feasible for an AI, and they can be done very exactly, but only for small versions of poker, for example where there are only a few cards in a pack. It’s very hard for an AI to do these calculations accurately for the full game of poker. As a result, over the decade or so that people have been working on scaling up poker, we’ve gradually seen improvements in the accuracy and efficiency of how to calculate these probabilities for larger and larger versions of poker.

所以，Libratus 是另一个令人印象深刻的现代人工智能应用。但考虑到扑克从一个版本到另一个稍大版本的转变花了十年时间，我不确定这些技术是否具有可扩展性。我认为还有一个合理的问题，即扑克中的博弈论思想在多大程度上可以延伸到现实世界。我们在日常生活中没有意识到有多少随机化，尽管这个世界肯定充满了代理；所以它应该是博弈论的，但我们在日常生活中并没有意识到多少随机化。

So yes, Libratus is another impressive modern AI application. But whether the techniques are at all scalable, given that it has taken a decade to go from one version of poker to another slightly larger version of poker, I’m not convinced. I think there’s also a reasonable question about how much those game-theoretic ideas in poker extend into the real world. We’re not aware of doing much randomization in our normal day-to-day lives, even though—for sure—the world is full of agents; so it ought to be game-theoretic, and yet we’re not aware of randomizing very much in our day-to-day lives.

马丁·福特：自动驾驶汽车是人工智能最受关注的应用之一。您认为什么时候全自动驾驶汽车才能成为真正实用的技术？想象一下，你在曼哈顿的某个地方，你叫了一辆 Uber，车上没有人，它会把你送到你指定的另一个地方。你认为，现实情况是多远？

MARTIN FORD: Self-driving cars are one of the highest-profile applications of AI. What is your estimate for when fully autonomous vehicles will become a truly practical technology? Imagine you’re in a random place in Manhattan, and you call up an Uber, and it’s going to arrive with no one in it, and then it will take you to another random place that you specify. How far off is that realistically, do you think?

斯图尔特·J·拉塞尔：是的，自动驾驶汽车的时间表是一个具体的问题，同时也是一个具有重要经济意义的问题，因为各大公司都在这些项目上投入大量资金。

STUART J. RUSSELL: Yes, the timeline for self-driving cars is a concrete question, and it’s also an economically important question because companies are investing a great deal in these projects.

值得注意的是，第一辆真正意义上的自动驾驶汽车在公共道路上行驶是在 30 年前！那是 Ernst Dickmanns 在德国进行的一次演示，演示了一辆汽车在高速公路上行驶、变道和超车。当然，困难在于信任：虽然你可以在短时间内成功进行演示，但你需要一个 AI 系统运行数十年而没有出现重大故障，才能成为一辆安全的汽车。

It is worth noting that the first actual self-driving car, operating on a public road, was 30 years ago! That was Ernst Dickmanns’ demo in Germany of a car driving on the freeway, changing lanes, and overtaking other vehicles. The difficulty of course is trust: while you can run a successful demonstration for a short time, you need an AI system to run for decades with no significant failures in order to qualify as a safe vehicle.

那么，挑战就在于建立一个人们愿意信任自己和孩子的人工智能系统，但我认为我们还没有做到这一点。

The challenge, then, is to build an AI system that people are willing to trust themselves and their kids to, and I don’t think we’re quite there.

目前在加州测试的车辆结果表明，人类仍然觉得他们必须像每英里道路测试那样频繁地进行干预。有更成功的人工智能驾驶项目，比如谷歌子公司 Waymo，他们取得了一些令人尊敬的记录；但我认为，他们还需要几年的时间才能在各种条件下做到这一点。

Results from vehicles that are being tested in California at the moment indicate that humans still feel they must intervene as frequently as once every mile of road testing. There are more successful AI driving projects, such as Waymo, which is the Google subsidiary working on this, that have some respectable records; but they are still, I think, several years away from being able to do this in a wide range of conditions.

大多数测试都是在路况良好、标记清晰的道路上进行的。众所周知，如果你在夜间开车，下着倾盆大雨，路面上会反射灯光，还可能有道路施工，车道标记可能会移动，等等……如果你按照原来的车道标记行驶，你现在就会直接撞到墙上了。我认为在这种情况下，人工智能系统真的很难应对。这就是为什么我认为，如果自动驾驶汽车问题在未来五年内得到充分解决，我们就很幸运了。

Most of these tests have been conducted in good conditions on well-marked roads. And as you know, when you’re driving at night and it’s pouring with rain, and there are lights reflecting off the road, and there may also be roadworks, and they might have moved the lane markers, and so on ... if you had followed the old lane markers, you’d have driven straight into a wall by now. I think in those kinds of circumstances, it’s really hard for AI systems. That’s why I think that we’ll be lucky if the self-driving car problem is solved sufficiently in the next five years.

当然，我不知道各大汽车公司有多少耐心。我确实认为每个人都坚信人工智能汽车终将到来，而各大汽车公司当然觉得他们必须早点到来，否则就会错过重大机遇。

Of course, I don’t know how much patience the major car companies have. I do think everyone is committed to the idea that AI-driven cars are going to come, and of course the major car companies feel they must be there early or miss a major opportunity.

马丁·福特：当人们问我关于自动驾驶汽车的问题时，我通常会告诉他们 10-15 年的时间范围。你估计的五年似乎相当乐观。

MARTIN FORD: I usually tell people a 10-15-year time frame when they ask me about self-driving cars. Your estimate of five years seems quite optimistic.

斯图尔特·J·拉塞尔：是的，五年是乐观的。正如我所说，我认为如果我们能在五年内看到无人驾驶汽车，那我们就很幸运了，而且这个时间可能更长。不过，有一点很清楚，随着我们积累了更多的经验，许多早期的无人驾驶汽车相当简单的架构理念现在都被抛弃了。

STUART J. RUSSELL: Yes, five years is optimistic. As I said, I think we’ll be lucky if we see driverless cars in five years, and it could well be longer. One thing that is clear, though, is that many of the early ideas of fairly simple architectures for driverless cars are now being abandoned, as we gain more experience.

在早期版本的谷歌汽车中，他们拥有基于芯片的视觉系统，能够很好地检测其他车辆、车道标记、障碍物和行人。这些视觉系统以某种逻辑形式有效地传递此类信息，然后控制器应用逻辑规则告诉汽车该做什么。问题是，谷歌每天都在添加新规则。也许他们会进入一个交通环岛（或我们在英国称之为环形交叉路口），然后会有一个小女孩在交通环岛处逆行。他们没有针对这种情况的规则。所以，他们必须添加一条新规则，依此类推。我认为这种架构从长远来看可能永远不可能奏效，因为总是有更多的规则需要编码，如果缺少某条规则，就可能事关道路上的生死。

In the early versions of Google’s car, they had chip-based vision systems that were pretty good at detecting other vehicles, lane markers, obstacles, and pedestrians. Those vision systems passed that kind of information effectively in a sort of logical form and then the controller applied logical rules telling the car what to do. The problem was that every day, Google found themselves adding new rules. Perhaps they would go into a traffic circle—or a roundabout, as we call them in England—and there would be a little girl riding her bicycle the wrong way around the traffic circle. They didn’t have a rule for that circumstance. So, then they have to add a new one, and so on, and so on. I think that there is probably no possibility that this type of architecture is ever going to work in the long run, because there are always more rules that should be encoded, and it can be a matter of life and death on the road if a particular rule is missing.

相比之下，我们下棋或围棋时不会制定一堆针对某一特定位置的规则——例如，如果某人的王在这里，车在那里，后在那里，那么就走这一步。我们编写国际象棋程序的方式不是这样的。我们编写国际象棋程序的方式是了解国际象棋规则，然后研究各种可能动作的后果。

By contrast, we don’t play chess or Go by having a bunch of rules specific to one exact position or another—for instance, saying if the person’s king is here and their rook is there, and their queen is there, then make this move. That’s not how we write chess programs. We write chess programs by knowing the rules of chess and then examining the consequences of various possible actions.

自动驾驶汽车人工智能必须以同样的方式处理道路上的意外情况，而不是通过特殊规则。当它没有针对当前情况的现成策略时，它应该使用这种基于前瞻的决策形式。如果人工智能没有这种方法作为后备，那么它在某些情况下就会出现失误，无法安全驾驶。当然，这在现实世界中是不够好的。

A self-driving car AI must deal with unexpected circumstances on the road in the same way, not through special rules. It should use this form of lookahead-based decision-making when it doesn’t have a ready-made policy for how to operate in the current circumstance. If an AI doesn’t have this approach as a fallback, then it’s going to fall through the cracks in some situations and fail to drive safely. That’s not good enough in the real world, of course.

马丁·福特：您已经指出了当前狭义或专业化人工智能技术的局限性。让我们来谈谈通用人工智能的前景，它有望在未来解决这些问题。您能准确解释一下什么是通用人工智能吗？通用人工智能的真正含义是什么？在实现通用人工智能之前，我们需要克服的主要障碍是什么？

MARTIN FORD: You’ve noted the limitations in current narrow or specialized AI technology. Let’s talk about the prospects for AGI, which promises to someday solve these problems. Can you explain exactly what Artificial General Intelligence is? What does AGI really mean, and what are the main hurdles we need to overcome before we can achieve AGI?

斯图尔特·J·拉塞尔：通用人工智能是一个最近才出现的术语，它实际上只是提醒我们在人工智能方面的真正目标——一种与我们自己的智能非常相似的通用智能。从这个意义上讲，通用人工智能实际上就是我们一直所说的人工智能。我们还没有完成，我们还没有创造出通用人工智能。

STUART J. RUSSELL: Artificial General Intelligence is a recently coined term, and it really is just a reminder of our real goals in AI—a general-purpose intelligence much like our own. In that sense, AGI is actually what we’ve always called artificial intelligence. We’re just not finished yet, and we have not created AGI yet.

人工智能的目标一直是创造通用智能机器。AGI 也提醒我们，我们人工智能目标中的“通用”部分经常被忽视，而更受关注的是更具体的子任务和应用任务。这是因为到目前为止，解决现实世界中的子任务（比如下棋）更容易。如果我们再看一眼 AlphaZero，它通常在双人确定性完全可观察棋盘游戏类中发挥作用。然而，它不是一种可以解决所有问题类别的通用算法。AlphaZero 无法处理部分可观察性；它无法处理不可预测性；并且它假设规则是已知的。可以说，AlphaZero 无法处理未知的物理现象。

The goal of AI has always been to create general-purpose intelligent machines. AGI is also a reminder that the “general-purpose” part of our AI goals has often been neglected in favor of more specific subtasks and application tasks. This is because it’s been easier so far to solve subtasks in the real world, such as playing chess. If we look again at AlphaZero for a moment, it generally works within the class of two-player deterministic fully-observable board games. However, it is not a general algorithm that can work across all classes of problems. AlphaZero can’t handle partial observability; it can’t handle unpredictability; and it assumes that the rules are known. AlphaZero can’t handle unknown physics, as it were.

现在，如果我们能够逐步消除 AlphaZero 的这些限制，我们最终将拥有一个能够在几乎任何情况下成功运作的人工智能系统。我们可以要求它设计一种新的高速船，或者为晚餐摆好桌子。我们可以要求它找出我们的狗出了什么问题，它应该能够做到这一点——甚至可能通过阅读所有已知的犬科医学知识，并利用这些信息找出我们的狗出了什么问题。

Now if we could gradually remove those limitations around AlphaZero, we’d eventually have an AI system that could learn to operate successfully in pretty much any circumstance. We could ask it to design a new high-speed watercraft, or to lay the table for dinner. We could ask it to figure out what’s wrong with our dog and it should be able to do that—perhaps even by reading everything about canine medicine that’s ever been known and using that information to figure out what’s wrong with our dog.

这种能力被认为反映了人类表现出的普遍性。原则上，只要有足够的时间，人类也可以做所有这些事情，甚至更多。这就是我们在谈论 AGI 时想到的普遍性概念：一种真正的通用人工智能。

This kind of capability is thought to reflect the generality of intelligence that humans exhibit. And in principle a human being, given enough time, could also do all of those things, and so very much more. That is the notion of generality that we have in mind when we talk about AGI: a truly general-purpose artificial intelligence.

当然，AGI 可能还能够做到人类做不到的其他事情。我们无法在头脑中计算百万位数的乘法，而计算机却可以相对轻松地做到这一点。因此，我们假设，事实上，机器可能能够表现出比人类更大的通用性。

Of course, there may be other things that humans can’t do that an AGI will be able to do. We can’t multiply million-digit numbers in our heads, and computers can do that relatively easily. So, we assume that in fact, machines may be able to exhibit greater generality than humans do.

然而，还值得指出的是，机器在以下意义上与人类相媲美的可能性非常小。一旦机器能够阅读，那么它基本上可以阅读所有已写成的书籍；而人类甚至无法阅读所有已写成的书籍中的一小部分。因此，一旦 AGI 的阅读水平超过幼儿园水平，它将超越人类曾经做过的任何事，并且它将拥有比人类更大的知识库。

However, it’s also worth pointing out that it’s very unlikely that there will ever be a point where machines are comparable to human beings in the following sense. As soon as machines can read, then a machine can basically read all the books ever written; and no human can read even a tiny fraction of all the books that have ever been written. Therefore, once an AGI gets past kindergarten reading level, it will shoot beyond anything that any human being has ever done, and it will have a much bigger knowledge base than any human ever has.

因此，从这个意义上以及其他许多意义上来说，很可能会发生的事情是，机器在各个重要方面的能力将远远超过人类。在其他方面，它们可能还相当落后，因此从这个意义上来说，它们看起来不像人类。但这并不意味着人类和 AGI 机器之间的比较毫无意义：从长远来看，重要的是我们与机器的关系，以及 AGI 机器在我们的世界中运行的能力。

And so, in that sense and many other senses, what’s likely to happen is that machines will far exceed human capabilities along various important dimensions. There may be other dimensions along which they’re fairly stunted and so they’re not going to look like humans in that sense. This doesn’t mean that a comparison between humans and AGI machines is meaningless though: what will matter in the long run is our relationship with machines, and the ability of the AGI machine to operate in our world.

在智力的某些方面（例如短期记忆），人类确实比猿类强；但尽管如此，毫无疑问哪个物种占主导地位。如果你是大猩猩或黑猩猩，你的未来完全掌握在人类手中。这是因为，尽管与大猩猩和猿类相比，我们的短期记忆相当可怜，但我们能够凭借在现实世界中的决策能力主宰它们。

There are dimensions of intelligence (for example, short-term memory) where humans are actually exceeded by apes; but nonetheless, there’s no doubt which of the species is dominant. And if you are a gorilla or a chimpanzee, your future is entirely in the hands of humans. Now that is because, despite our fairly pathetic short-term memories compared to gorillas and apes, we are able to dominate them because of our decision-making capabilities in the real world.

当我们创造 AGI 时，我们无疑将面临同样的问题：如何避免大猩猩和黑猩猩的命运，而不是将我们自己未来的控制权交给 AGI。

We will undoubtedly face this same issue when we create AGI: how to avoid the fate of the gorilla and the chimpanzee, and not cede control of our own future to that AGI.

马丁·福特：这个问题很可怕。之前，您谈到人工智能的概念突破往往比现实提前几十年。您是否看到任何迹象表明，创建 AGI 的概念突破已经实现，还是 AGI 还遥遥无期？

MARTIN FORD: That’s a scary question. Earlier, you talked about how conceptual breakthroughs in AI often run decades ahead of reality. Do you see any indications that the conceptual breakthroughs for creating AGI have already been made, or is AGI still far in the future?

STUART J. RUSSELL：我确实觉得 AGI 的许多概念构建模块已经存在了，是的。我们可以开始探索这个问题，先问自己：“为什么深度学习系统不能成为 AGI 的基础，它们有什么问题？”

STUART J. RUSSELL: I do feel that many of the conceptual building blocks towards AGI pieces are already here, yes. We can start to explore this question by asking ourselves: “Why can’t deep learning systems be the basis for AGI, what’s wrong with them?”

很多人可能会这样回答我们的问题：“深度学习系统很好，但我们不知道如何存储知识，如何推理，如何构建更具表现力的模型，因为深度学习系统只是电路，而电路的表达能力毕竟不强。”

A lot of people might answer our question by saying: “Deep learning systems are fine, but we don’t know how to store knowledge, or how to do reasoning, or how to build more expressive kinds of models, because deep learning systems are just circuits, and circuits are not very expressive after all.”

当然，正是因为电路的表达能力不强，所以没人会考虑用电路来编写工资软件。我们改用编程语言来创建工资软件。用电路编写的工资软件将长达数十亿页，而且完全无用且不灵活。相比之下，编程语言的表达能力非常强，功能非常强大。事实上，它们是表达算法过程的最强大的东西。

And for sure, it’s because circuits are not very expressive that no one thinks about writing payroll software using circuits. We instead use programming languages to create payroll software. Payroll software written using circuits would be billions of pages long and completely useless and inflexible. By comparison, programming languages are very expressive and very powerful. In fact, they are the most powerful things that can exist for expressing algorithmic processes.

事实上，我们已经知道如何表示知识以及如何进行推理：我们已经开发了计算逻辑很长时间了。甚至在计算机出现之前，人们就在思考进行逻辑推理的算法程序。

In fact, we already know how to represent knowledge and how to do reasoning: we have developed computational logic over quite a long time now. Even predating computers, people were thinking about algorithmic procedures for doing logical reasoning.

因此，可以说，AGI 的一些概念性构建模块已经存在了几十年。我们只是还没有弄清楚如何将它们与深度学习的惊人学习能力结合起来。

And so, arguably, some of the conceptual building blocks for AGI have already been here for decades. We just haven’t figured out yet how to combine those with the very impressive learning capacities of deep learning.

人类还已经建立了一种称为概率编程的技术，我认为它确实将学习能力与逻辑语言和编程语言的表达能力结合起来。从数学上讲，这种概率编程系统是一种写下概率模型的方法，然后将其与证据结合起来，使用概率推理来产生预测。

The human race has also already built a technology called probabilistic programming, which I will say does combine learning capabilities with the expressive power of logical languages and programming languages. Mathematically speaking, such a probabilistic programming system is a way of writing down probability models which can then be combined with evidence, using probabilistic inference to produce predictions.

在我的团队中，我们有一种语言叫做 BLOG，即贝叶斯逻辑。BLOG 是一种概率建模语言，因此你可以将你所知道的内容以 BLOG 模型的形式写下来。然后，你将这些知识与数据相结合，并进行推理，从而做出预测。

In my group we have a language called BLOG, which stands for Bayesian Logic. BLOG is a probabilistic modeling language, so you can write down what you know in the form of a BLOG model. You then combine that knowledge with data, and you run inference, which in turn makes predictions.

这种系统的一个真实示例是核试验禁令条约的监测系统。它的工作方式是，我们写下我们所知道的地球物理知识，包括地震信号在地球中的传播、地震信号的检测、噪声的存在、检测站的位置等等。这就是模型——它以形式语言表达，连同所有不确定性：例如，我们预测信号在地球中传播速度的能力的不确定性。数据是来自遍布世界各地的检测站的原始地震信息。然后是预测：今天发生了哪些地震事件？它们发生在哪里？它们有多深？它们有多大？也许：哪些可能是核爆炸？该系统是当今核试验禁令条约的主动监测系统，它似乎运行良好。

A real-world example of such a system is the monitoring system for the nuclear test-ban treaty. The way it works is that we write down what we know about the geophysics of the earth, including the propagation of seismic signals through the earth, the detection of seismic signals, the presence of noise, the locations of detection stations, and so on. That’s the model—which is expressed in a formal language, along with all the uncertainties: for example, uncertainty in our ability to predict the speed of propagation of a signal through the earth. The data is the raw seismic information coming from the detection stations that are scattered around the world. Then there is the prediction: What seismic events took place today? Where did they take place? How deep were they? How big were they? And perhaps: Which ones are likely to be nuclear explosions? This system is an active monitoring system today for the test-ban treaty, and it seems to be working pretty well.

总而言之，我认为 AGI 或人类智能所需的许多概念构建模块已经存在。但仍有一些缺失的部分。其中之一就是如何理解自然语言以产生推理过程可以操作的知识结构。典型的例子可能是：AGI 如何阅读化学教科书，然后解决一堆化学考试问题（不是多项选择题，而是真正的化学考试问题），并出于正确的理由解决这些问题，展示得出答案的推导和论据？然后，如果以一种优雅而有原则的方式完成，AGI 应该能够阅读物理教科书、生物教科书和材料教科书等等。

So, to summarize, I think that many of the conceptual building blocks needed for AGI or human-level intelligence are already here. But there are some missing pieces. One of them is a clear approach to how natural language can be understood to produce knowledge structures upon which reasoning processes can operate. The canonical example might be: How can an AGI read a chemistry textbook and then solve a bunch of chemistry exam problems—not multiple choice but real chemistry exam problems—and solve them for the right reasons, demonstrating the derivations and the arguments that produced the answers? And then, presumably if that’s done in a way that’s elegant and principled, the AGI should then be able to read a physics textbook and a biology textbook and a materials textbook, and so on.

马丁·福特：或者我们可以想象一个 AGI 系统从历史书中获取知识，然后将所学知识应用于当代地缘政治的模拟，或者类似的东西，它实际上是在转移知识并将其应用于完全不同的领域？

MARTIN FORD: Or we might imagine an AGI system acquiring knowledge from, say, a history book and then applying what it’s learned to a simulation of contemporary geopolitics, or something like that, where it’s really moving knowledge and applying it in an entirely different domain?

斯图尔特·J·拉塞尔：是的，我认为这是一个很好的例子，因为它与人工智能系统在地缘政治或金融方面操纵现实世界的能力有关。

STUART J. RUSSELL: Yes, I think that’s a good example because it relates to the ability of an AI system to then be able to manipulate the real world in a geopolitical sense or a financial sense.

例如，如果人工智能为首席执行官提供公司战略建议，它可能能够通过设计一些令人惊叹的产品营销收购策略有效地超越所有其他公司。

If, for example, the AI is advising a CEO on corporate strategy, it might be able to effectively outplay all the other companies by devising some amazing product marketing acquisition strategies, and so on.

所以，我认为理解语言的能力，并根据理解的结果进行操作，是 AGI 尚未实现的一个重要突破。

So, I’d say that the ability to understand language, and then to operate with the results of that understanding, is one important breakthrough for AGI that still needs to happen.

另一项尚未实现的 AGI 突破是能够长时间运行。虽然 AlphaZero 是一个非常棒的问题解决系统，可以思考未来 20 步甚至 30 步，但与人类大脑每时每刻所做的相比，这仍然微不足道。人类在原始步骤中使用我们发送给肌肉的运动控制信号；而仅仅输入一段文字就需要数千万个运动控制命令。因此，AlphaZero 的这 20 或 30 步只能在未来几毫秒内获得 AGI。正如我们之前所讨论的，AlphaZero 对于规划机器人的活动完全没用。

Another AGI breakthrough still to happen is the ability to operate over long timescales. While AlphaZero is an amazingly good problem-solving system which can think 20, sometimes 30 steps into the future, that is still nothing compared to what the human brain does every moment. Humans, in our primitive steps, use motor control signals that we send to our muscles; and just typing a paragraph of text is several tens of millions of motor control commands. So those 20 or 30 steps by AlphaZero would only get an AGI only a few milliseconds into the future. As we talked about earlier, AlphaZero would be totally useless for planning the activity of a robot.

马丁·福特：人类在环游世界时需要进行如此多的计算和决策，如何解决这个问题呢？

MARTIN FORD: How do humans even solve this problem with so many calculations and decisions to be made as they navigate the world?

斯图尔特·J·拉塞尔：人类和机器人在现实世界中运作的唯一方式是在多个抽象尺度上运作。我们不会根据具体要按照什么顺序去做某件事来规划我们的生活。相反，我们会这样规划我们的生活：“好吧，今天下午我要试着写另一章书”，然后：“这本书会讲这样那样的事情。”或者像“明天我要坐飞机飞回巴黎。”

STUART J. RUSSELL: The only way that humans and robots can operate in the real world is to operate at multiple scales of abstraction. We don’t plan our lives in terms of exactly which thing are we going to actuate in exactly which order. We instead plan our lives in terms of “OK, this afternoon I’m going try to write another chapter of my book” and then: “It’s going to be about such and such.” Or things like, “Tomorrow I’m going to get on the plane and fly back to Paris.”

这些都是我们的抽象行为。然后，当我们开始更详细地规划它们时，我们会将它们分解成更精细的步骤。这是人类的常识。我们一直都在这样做，但我们实际上并不太了解如何让人工智能系统做到这一点。特别是，我们还不了解如何让人工智能系统首先构建这些高级动作。行为肯定是按层次结构组织到这些抽象层中的，但层次结构从何而来？我们如何创建它然后使用它？

Those are our abstract actions. And then as we start to plan them in more detail, we break them down into finer steps. That’s common sense for humans. We do this all the time, but we actually don’t understand very well how to have AI systems do this. In particular, we don’t understand yet how to have AI systems construct those high-level actions in the first place. Behavior is surely organized hierarchically into these layers of abstraction, but where does the hierarchy come from? How do we create it and then use it?

如果我们能够解决人工智能的这个问题，如果机器可以开始构建自己的行为层次结构，使它们能够在长期内成功地在复杂的环境中运行，这将是 AGI 的一大突破，使我们在现实世界中距离实现人类水平的功能还有很长的路要走。

If we can solve this problem for AI, if machines can start to construct their own behavioral hierarchies that allow them to operate successfully in complex environments over long timescales, that will be a huge breakthrough for AGI that takes us a long way towards a human-level functionality in the real world.

马丁·福特：您预测我们何时能够实现 AGI？

MARTIN FORD: What is your prediction for when we might achieve AGI?

斯图尔特·J·拉塞尔：这些突破与更大的数据集或更快的机器无关，因此我们无法对它们何时会发生做出任何定量预测。

STUART J. RUSSELL: These kinds of breakthroughs have nothing to do with bigger datasets or faster machines, and so we can’t make any kind of quantitative prediction about when they’re going to occur.

我总是讲述核物理学中发生的事情。1933 年 9 月 11 日，欧内斯特·卢瑟福表达的共识是，从原子中提取原子能是永远不可能的。因此，他的预测是“永远不可能”，但事实是，第二天早上，利奥·西拉德读了卢瑟福的演讲，对此感到恼火，并发明了一种由中子介导的核链式反应！卢瑟福的预测是“永远不可能”，真相大约在 16 小时后出现。同样，我觉得很难对这些 AGI 的突破何时到来做出定量预测，但卢瑟福的故事是一个好故事。

I always tell the story of what happened in nuclear physics. The consensus view as expressed by Ernest Rutherford on September 11th, 1933, was that it would never be possible to extract atomic energy from atoms. So, his prediction was “never”, but what turned out to be the case was that the next morning Leo Szilard read Rutherford’s speech, became annoyed by it, and invented a nuclear chain reaction mediated by neutrons! Rutherford’s prediction was “never” and the truth was about 16 hours later. In a similar way, it feels quite futile for me to make a quantitative prediction about when these breakthroughs in AGI will arrive, but Rutherford’s story is a good one.

马丁·福特：您认为通用人工智能会在您有生之年出现吗？

MARTIN FORD: Do you expect AGI to happen in your lifetime?

斯图尔特·J·拉塞尔：当被问及时，我有时会说是的，我预计 AGI 会在我孩子的有生之年实现。当然，这是我的含糊其辞，因为到那时我们可能会有一些延长寿命的技术，所以这可能会延长相当长的时间。

STUART J. RUSSELL: When pressed, I will sometimes say yes, I expect AGI to happen in my children’s lifetime. Of course, that’s me hedging a bit because we may have some life extension technologies in place by then, so that could stretch it out quite a bit.

但鉴于我们对这些突破有足够的了解，至少可以描述它们，而且人们肯定对它们的解决方案有所了解，在我看来，我们只是在等待一点灵感。

But given that we kind of understand enough about these breakthroughs to at least describe them, and that people certainly have inklings of what their solutions might be, suggests to me that we’re just waiting for a bit of inspiration.

此外，许多非常聪明的人正在研究这些问题，可能比该领域历史上任何时候都多，这主要归功于谷歌、Facebook、百度等。现在，大量资源被投入到人工智能中。学生对人工智能也非常感兴趣，因为它现在非常令人兴奋。

Furthermore, a lot of very smart people are working on these problems, probably more than ever in the history of the field, mainly because of Google, Facebook, Baidu, and so on. Enormous resources are being put into AI now. There’s also enormous student interest in AI because it’s so exciting right now.

因此，所有这些因素都使人们相信，突破发生的速度可能相当高。这些突破的规模肯定与人工智能过去 60 年发生的十几个概念突破相当。

So, all those things lead one to believe that the rate of breakthroughs occurring is probably likely to be quite high. These breakthroughs are certainly comparable in magnitude to a dozen of the conceptual breakthroughs that happened over the last 60 years of AI.

所以这就是为什么大多数人工智能研究人员都认为 AGI 是在不远的将来才会出现。它不是几千年后才会出现，甚至可能也不是几百年后才会出现。

So that is why most AI researchers have a feeling that AGI is something in the not-too-distant future. It’s not thousands of years in the future, and it’s probably not even hundreds of years in the future.

马丁·福特：您认为当第一个 AGI 被创造出来时会发生什么？

MARTIN FORD: What do you think will happen when the first AGI is created?

斯图尔特·J·拉塞尔：当这一切发生时，我们跨越的不仅仅是一条终点线。它将沿着多个维度发展。我们将看到机器超越人类的能力，就像它们在算术、国际象棋、围棋和视频游戏中所做的那样。我们将看到智能的其他各个维度和问题类别相继出现；这些将对人工智能系统在现实世界中的作用产生影响。例如，AGI 系统可能拥有超越人类的战略推理工具，我们将其用于军事和企业战略等。但这些工具可能先于阅读和理解复杂文本的能力。

STUART J. RUSSELL: When it happens, it’s not going to be a single finishing line that we cross. It’s going to be along several dimensions. We’ll see machines exceeding human capacities, just as they have in arithmetic, and now chess, Go, and in video games. We’ll see various other dimensions of intelligence and classes of problems that fall, one after the other; and those will then have implications for what AI systems can do in the real world. AGI systems may, for example, have strategic reasoning tools that are superhuman, and we use those for military and corporate strategy, and so on. But those tools may precede the ability to read and understand complex text.

早期的 AGI 系统本身仍然无法了解世界运作的一切，也无法控制这个世界。

An early AGI system, by itself, still won’t be able to learn everything about how the world works or be able to control that world.

我们仍需要为早期的 AGI 系统提供大量知识。不过，这些 AGI 看起来不会像人类，它们的能力甚至不会与人类大致相同。这些 AGI 系统在不同方向上会非常尖锐。

We’ll still need to provide a lot of the knowledge to those early AGI systems. These AGIs are not going to look like humans though, and they won’t have even roughly the same abilities across even roughly the same spectrum as humans. These AGI systems are going to be very spiky in different directions.

马丁·福特：我想进一步谈谈与人工智能和通用人工智能相关的风险。我知道这是你最近工作的重点。

MARTIN FORD: I want to talk more about the risks associated with AI and AGI. I know that’s an important focus of your recent work.

我们先从人工智能的经济风险开始，当然，我在上一本书《机器人崛起》中写过这个话题。很多人认为，我们正处于新工业革命的前沿。这场革命将彻底改变就业市场、经济等。您对此持什么看法？这种说法是否被夸大了，还是您认同这种说法？

Let’s start with the economic risks of AI, which is the thing that, of course, I’ve written about in my previous book, Rise of the Robots. A lot of people believe that we are on the leading edge of something on the scale of a new industrial revolution. Something that’s going to be totally transformative in terms of the job market, the economy and so forth. Where do you fall on that? Is that overhyped, or would you line up with that assertion?

斯图尔特·J·拉塞尔：我们讨论过，人工智能和通用人工智能的突破时间表很难预测。这些突破将使人工智能能够完成人类目前所做的许多工作。同样很难预测哪些就业类别的顺序将面临机器替代的风险，以及围绕这一风险的时间表。

STUART J. RUSSELL: We’ve discussed how the timeline for breakthroughs in AI and AGI is hard to predict. Those are the breakthroughs that will enable an AI to do a lot of the jobs that humans do right now. It’s also quite hard to forecast which sequence of employment categories are going to be at risk from machine replacement and a timeline around that.

然而，我从许多讨论和演讲中看到，人们可能高估了当前的人工智能技术所能做的事情，而且将我们所知的技术融入现有的企业和政府等极其复杂的功能中也存在困难。

However, what I see in a lot of the discussions and presentations from people talking about this, is that there’s probably an over-estimate of what current AI technologies are able to do and also, the difficulty of integrating what we know how to do into the existing extremely complex functionality of corporations and governments, and so on.

我确实同意，过去几百年来存在的许多工作都是重复性的，从事这些工作的人基本上是可以互换的。如果这是一项需要你雇佣数百人或数千人来做的工作，而且你可以确定这些人所做的是重复性的特定任务，那么这类工作就很容易受到影响。这是因为你可以说，在这些工作中，我们将人类当作机器人来使用。因此，当我们有了真正的机器人，它们将能够完成这些工作，这并不奇怪。

I do agree that a lot of jobs that have existed for the last few hundred years are repetitive, and the humans who are doing them are basically exchangeable. If it’s a job where you hire people by the hundred or by the thousand to do it, and you can identify what that person does as a particular task that is then repeated over and over again, those kinds of jobs are going to be susceptible. That’s because you could say that, in those jobs, we are using humans as robots. So, it’s not surprising that when we have real robots, they’re going to be able to do those jobs.

我还认为，目前各国政府的心态是：“哦，那好吧。我想我们真的需要开始培训人们成为数据科学家，因为这是未来的工作——或者机器人工程师。”这显然不是解决方案，因为我们不需要十亿数据科学家和机器人工程师：我们只需要几百万人。这可能是新加坡这样的小国的战略；或者在我目前所在的迪拜，这也可能是一个可行的战略。但对于任何大国来说，这都不是一个可行的战略，因为这些地区根本没有足够的就业机会。这并不是说现在没有工作：肯定有，培训更多人来做这些工作是有意义的；但这根本不能解决长期问题。

I also think that the current mindset among governments is: “Oh, well then. I guess we really need to start training people to be data scientists, because that’s the job of the future—or robot engineers.” This clearly isn’t the solution because we don’t need a billion data scientists and robot engineers: we just need a few million. This might be a strategy for a small country like Singapore; or where I am currently, in Dubai, it might also be a viable strategy. But it’s not a viable strategy for any major country because there is simply not going to be enough jobs in those areas. That’s not to say that there are no jobs now: there certainly are, and training more people to do them makes sense; but this simply is not a solution to the long-term problem.

从长远来看，我认为人类经济的未来实际上只有两种。

There are really only two futures for the human economy that I see in the long run.

首先，实际上，大多数人并没有做任何被认为具有经济效益的事情。他们没有以任何形式参与经济劳动换取报酬，而这正是全民基本收入的愿景：经济中有一个部门基本上是自动化的，而且生产力极高，生产力以商品和服务的形式创造财富，而这些财富最终以某种方式补贴了其他所有人的经济生存能力。在我看来，这似乎不是一个非常有趣的世界，至少就其本身而言，没有很多其他东西来让生活变得有价值，并为人们提供足够的激励去做我们现在所做的所有事情。例如，上学、学习和培训，以及成为各个领域的专家。如果教育没有任何经济功能，那么很难看到接受良好教育的动机。

The first is that effectively, most people are not doing anything that’s considered economically productive. They’re not involved in economic exchange of work for pay in any form, and this is the vision of the universal basic income: that there is a sector of the economy that is largely automated and incredibly productive, and that productivity generates wealth, in the form of goods and services, that in one way or another ends up subsidizing the economic viability of everyone else. That to me does not seem like a very interesting world to live in, at least not by itself, without a lot of other things needed to go on to make life worth living and provide sufficient incentive for people to do all of the things that we do now. For example, going to school, learning and training, and becoming experts in various areas. It’s hard to see the motivation for acquiring a good education when it doesn’t have any economic function.

从长远来看，我能预见的第二个未来是，尽管机器将完成许多商品和基本服务，如交通运输，但人类仍然可以做一些事情来改善自己和他人的生活质量。有人能够教导、激励人们过上更丰富、更有趣、更丰富多彩和更充实的生活，无论是教人们欣赏文学或音乐，如何建造，甚至如何在荒野中生存。

The second of the two futures I can see in the long run is that even though machines will be doing a lot of goods and basic services like transportation, there are still things that people can do which improve the quality of life for themselves and for others. There are people who are able to teach, to inspire people to live richer, more interesting, more varied and more fulfilling lives, whether that’s teaching people to appreciate literature or music, how to build, or even how to survive in the wilderness.

马丁·福特：您认为，一旦人工智能改变了我们的经济，我们作为个体和物种是否能够走向积极的未来？

MARTIN FORD: Do you think we can navigate as individuals and as a species towards a positive future, once AI has changed our economy?

斯图尔特·J·拉塞尔：是的，我确实这么认为，但我认为积极的未来需要人类的干预，以帮助人们过上积极的生活。我们现在需要开始积极地走向一个能够为人们带来最具建设性的挑战和最有趣的生活体验的未来。一个能够建立情感韧性并培养对自己和他人生活普遍建设性和积极态度的世界。目前，我们在这方面做得很糟糕。所以，我们现在必须开始改变这种状况。

STUART J. RUSSELL: Yes, I really do, but I think that a positive future will require human intervention to help people live positive lives. We need to start actively navigating, right now, towards a future that can present the most constructive challenges and the most interesting experiences in life for people. A world that can build emotional resilience and nurture a generally constructive and positive attitude to one’s own life—and to the lives of others. At the moment, we are pretty terrible at doing that. So, we have to start changing that now.

我认为，我们还需要从根本上改变我们对科学的用途和科学能为我们做什么的态度。我的口袋里有一部手机，人类可能在科学和工程上花费了数万亿美元，最终创造了像我的手机这样的产品。然而，我们几乎没有花任何钱去了解人们如何过上有趣而充实的生活，以及如何帮助我们周围的人做到这一点。我认为，作为一个种族，我们需要开始承认，如果我们以正确的方式帮助他人，这将为他们的余生创造巨大的价值。目前，我们几乎没有科学基础来做到这一点，我们没有学位课程，我们几乎没有关于它的期刊，那些正在尝试的期刊也没有得到认真对待。

I think that we’ll also need to fundamentally change our attitude about what science is for and what it can do for us. I have a cell phone in my pocket, and the human race probably spent on the order of a trillion dollars on the science and engineering that went into ultimately creating things like my cell phone. And yet we spend almost nothing on understanding how people can live interesting and fulfilling lives, and how we can help people around us do that. I think as a race that we will need to start acknowledging that if we help another person in the right way, it creates enormous value for them for the rest of their lives. Right now, we have almost no science base for how to do this, we have no degree programs in how to do it, we have very few journals about it, and those that are trying are not taken very seriously.

未来可以拥有一个完美运转的经济，在那里，善于过好生活、帮助他人的人可以提供这些服务。这些服务可能是辅导、可能是教学、可能是安慰，也可能是合作，这样我们都能拥有一个美好的未来。

The future can have a perfectly functioning economy where people who are expert in living life well, and helping other people, can provide those kinds of services. Those services may be coaching, they may be teaching, they may be consoling, or maybe collaborating, so that we can all really have a fantastic future.

这根本不是一个暗淡的未来：它比我们现在所拥有的未来要好得多；但它需要我们重新思考我们的教育体系、我们的科学基础和我们的经济结构。

It’s not a grim future at all: it’s a far better future than what we have at present; but it requires rethinking our education system, our science base, our economic structures.

我们现在需要从经济角度了解这将如何影响未来的收入分配。我们希望避免出现这样的情况：超级富豪拥有生产资料——机器人和人工智能系统——然后是他们的仆人，而世界其他人则无所事事。从经济角度来看，这是最糟糕的结果。

We need now to understand how this will function from an economic point of view in terms of the future distribution of income. We want to avoid a situation where there are the super-rich who own the means of production—the robots and the AI systems—and then there are their servants, and then there is the rest of the world doing nothing. That’s sort of the worst possible outcome from an economic point of view.

因此，我确实认为，一旦人工智能改变了人类经济，就会有一个积极的未来，但我们现在需要更好地了解它将会是什么样子，这样我们才能制定实现这一目标的计划。

So, I do think that there is a positive future that makes sense once AI has changed the human economy, but we need to get a better handle on what that’s going to look like now, so that we can construct a plan for getting there.

马丁·福特：您曾在伯克利和附近的加州大学旧金山分校致力于将机器学习应用于医疗数据。您是否认为人工智能会通过医疗保健和医学的进步为人类创造更积极的未来？

MARTIN FORD: You’ve worked on applying machine learning to medical data at both Berkeley and nearby UCSF. Do you think artificial intelligence will create a more positive future for humans through advances in healthcare and medicine?

斯图尔特·J·拉塞尔：是的，我认为是这样，但我还认为，在医学领域，我们对人体生理学了解很多——因此对我来说，基于知识或基于模型的方法比数据驱动的机器学习系统更有可能成功。

STUART J. RUSSELL: I think so, yes, but I also think that medicine is an area where we know a great deal about human physiology—and so to me, knowledge-based or model-based approaches are more likely to succeed than data-driven machine learning systems.

我认为深度学习不会适用于许多重要的医疗应用。如今我们可以从数百万患者那里收集数 TB 的数据，然后将这些数据放入黑箱学习算法中，这种想法对我来说毫无意义。当然，在医学的某些领域，数据驱动的机器学习可能非常有效。基因组数据就是其中一个领域；预测人类对各种遗传相关疾病的易感性。此外，我认为深度学习人工智能将擅长预测特定药物的潜在功效。

I don’t think that deep learning is going to work for a lot of important medical applications. The idea that today we can just collect terabytes of data from millions of patients and then throw that data into a black-box learning algorithm, doesn’t make sense to me. There may be some areas of medicine where data-driven machine learning works very well of course. Genomic data is one area; and predicting human susceptibility to various kinds of genetically-related diseases. Also, I think, deep learning AI will be strong at predicting the potential efficacy of particular drugs.

但这些例子距离人工智能能够像医生一样行动并能够判断患者的脑室是否阻塞，从而干扰脑脊液循环还很远。要真正做到这一点，更像是诊断汽车的哪个部件出现故障。如果你不知道汽车是如何工作的，那么找出风扇皮带坏了就会非常非常困难。

But these examples are a long way from an AI being able to act like a doctor and being able to decide, perhaps, that a patient has a blocked ventricle in the brain that’s interfering with the circulation of cerebral spinal fluid. To really do that, is more like diagnosing which part of a car is not working. If you have no idea how cars work, then figuring out that it’s the fan belt that’s broken is going to be very, very difficult.

当然，如果你是汽车修理专家，知道汽车的工作原理，并且知道一些症状，比如汽车发出扑通扑通的声音，汽车过热，那么你通常可以很快找出原因。人体生理学也一样，只是在建立这些人体生理学模型时必须付出巨大努力。

Of course, if you’re an expert car mechanic and you know how it all works, and you’ve got some symptoms to work with, maybe there’s a kind of a flapping noise and that the car’s overheating, then you generally can figure it out quickly. And it’s going to be the same with human physiology, except that there is a significant effort that must be put in into building these models of human physiology.

20 世纪 60 年代和 70 年代，人们已经为这些模型付出了很多努力，它们在一定程度上帮助了医学领域的人工智能系统取得进步。但今天，我们拥有的技术可以特别代表这些模型中的不确定性。机械系统模型是确定性的，具有特定的参数值：它们恰好代表一个完全可预测的虚构人类。

A lot effort was already put in to these models in the ‘60s and ‘70s, and they have helped AI systems in medicine progress to some degree. But today we have technology that can in particular represent the uncertainty in those models. Mechanical systems models are deterministic and have specific parameter values: they represent exactly one completely predictable, fictitious human.

另一方面，今天的概率模型可以代表整个人口，它们可以准确地反映我们可能存在的不确定性程度，例如，无法准确预测某人何时会心脏病发作。从个人层面预测心脏病发作之类的事情非常困难，但我们可以预测每个人的概率都有一定的，在剧烈运动或压力下可能会增加，而且这个概率取决于个人的各种特征。

Today’s probabilistic models, on the other hand, can represent an entire population, and they can accurately reflect the degree of uncertainty we might have about being able to predict, for example, exactly when someone is going to have a heart attack. It’s very hard to predict things like heart attacks on an individual level, but we can predict that there’s a certain probability per person, which might be increased during extreme exercise or stress, and that this probability would depend on various characteristics of the individual.

这种更现代、更概率的方法比以前的系统表现得更合理。概率系统使我们能够将人类生理学的经典模型与观察和实时数据结合起来，做出强有力的诊断并制定治疗计划。

This more modern and probabilistic approach behaves much more reasonably than previous systems. Probabilistic systems enable us to combine the classical models of human physiology with observation and real-time data, to make strong diagnosis and plan treatments.

马丁·福特：我知道您非常关注武器化人工智能的潜在风险。您能详细谈谈这个问题吗？

MARTIN FORD: I know you’ve focused a lot on the potential risks of weaponized AI. Could you talk more about that?

斯图尔特·J·拉塞尔：是的，我认为自主武器正在创造新军备竞赛的前景。这场军备竞赛可能已经导致致命自主武器的发展。这些自主武器可以赋予一些武器能够自行完成的任务描述，例如识别、选择和攻击人类目标。

STUART J. RUSSELL: Yes, I think autonomous weapons are now creating the prospect of a new arms race. This arms race may already be leading towards the development of lethal autonomous weapons. These autonomous weapons can be given some mission description that the weapon has the ability to achieve by itself, such as identifying, selecting, and attacking human targets.

有道德观点认为，这将超越人工智能的基本界限：我们将生杀之权交给机器来决定，这将从根本上降低我们重视人类生命和人类生命尊严的态度。

There are moral arguments that this will cross a fundamental line for artificial intelligence: that we’re handing over the power over life and death to a machine to decide, and that is a fundamental reduction in the way we value human life and the dignity of human life.

但我认为更实际的论点是，自主性的逻辑结果是可扩展性。由于每个自主武器都不需要个人监督，因此有人可以发射任意数量的武器。有人可以发动袭击，控制室里的五个人可以发射 10,000,000 件武器，消灭某个国家 12 岁至 60 岁之间的所有男性。因此，这些可能是大规模杀伤性武器，它们具有可扩展性：有人可以用 10 件、1,000 件、1,000,000 件或 10,000,000 件武器发动袭击。

But I think a more practical argument is that a logical consequence of autonomy is scalability. Since no supervision is required by an individual human for each individual autonomous weapon, someone could launch as many weapons as they want. Someone can launch an attack, where five guys in a control room could launch 10,000,000 weapons and wipe out all males between the age of 12 and 60 in some country. So, these can be weapons of mass destruction, and they have this property of scalability: someone could launch an attack with 10, or 1,000, or 1,000,000 or 10,000,000 weapons.

核武器一旦被使用，就会有人跨过一个重大门槛，而我们迄今为止在竞争中侥幸避免了这一门槛。自 1945 年以来，我们一直设法避免跨过这一门槛。但自主武器没有这样的门槛，因此事态可以更顺利地升级。它们也很容易扩散，因此一旦大量生产，它们很可能就会出现在国际武器市场上，那些比西方国家更无所顾忌的人就会得到它们。

With nuclear weapons, if they were used at all, someone would be crossing a major threshold which we’ve managed to avoid so far as a race, by the skin of our teeth. We have managed to avoid crossing that threshold since 1945. But autonomous weapons don’t have such a threshold, and so things can more smoothly escalate. They are also easily proliferated, so once they are manufactured in very large numbers it’s quite likely they’ll be on the international arms market and they’ll be accessible to people who have less scruples than, you know, the Western powers.

马丁·福特：商业应用和潜在军事应用之间存在大量技术转让。你可以在亚马逊上购买可能被用作武器的无人机……

MARTIN FORD: There’s a lot of technology transfer between commercial applications and potential military applications. You can buy a drone on Amazon that could potentially be weaponized...

斯图尔特·J·拉塞尔：目前，你可以购买遥控无人机，也许具有第一人称视角。你当然可以在上面安装一枚小型炸弹，然后投掷它杀死某人，但那是遥控飞行器，这是不同的。它不可扩展，因为除非你拥有 1000 万名飞行员，否则你无法发射 1000 万架这样的无人机。所以，当然需要训练整个国家来做到这一点，或者他们也可以为这 1000 万人提供机枪，然后去杀人。值得庆幸的是，我们有一个国际制裁控制系统和军事准备系统等，试图防止这些事情发生。但我们没有一个可以对抗自主武器的国际控制系统。

STUART J. RUSSELL: So, at the moment, you can buy a drone that’s remotely piloted, maybe with first-person vision. You could certainly attach a little bomb to it and deliver it and kill someone, but that’s a remotely piloted vehicle, which is different. It’s not scalable because you can’t launch 10,000,000 of those unless you’ve got 10,000,000 pilots. So, someone would need a whole country trained to do that, of course, or they could also give those 10,000,000 people machine guns and then go and kill people. Thankfully we have an international system of control of sanctions, and military preparedness, and so on—to try to prevent these things from happening. But we don’t have an international system of control that would work against autonomous weapons.

马丁·福特：那么，难道几个人在地下室里就不能开发自己的自主控制系统，然后将其部署在商用无人机上吗？我们如何才能控制这些自制的人工智能武器？

MARTIN FORD: Still, couldn’t a few people in a basement somewhere develop their own autonomous control system and then deploy it on commercially available drones? How would we be able control those kinds of homemade AI weapons?

斯图尔特·J·拉塞尔：是的，可以设想部署类似于控制自动驾驶汽车的软件来控制投掷炸弹的四轴飞行器。这样，你就可能拥有了类似自制自主武器的东西。根据条约，可能会有一个核查机制，要求无人机制造商和自动驾驶汽车芯片制造商等合作，这样任何大批量订购的人都会受到注意——就像任何订购大量化学武器前体化学品的人都无法逃脱惩罚一样，因为根据化学武器条约，公司必须了解其客户，并报告任何试图大批量购买某些危险产品的行为。

STUART J. RUSSELL: Yes, something resembling the software that controls a self-driving car could conceivably be deployed to control a quadcopter that delivers a bomb. Then you might have something like a homemade autonomous weapon. It could be that under a treaty, there would be a verification mechanism that would require the cooperation of drone manufacturers and the people who make chips for self-driving cars and so on, so that anyone ordering large quantities would be noticed—in the same way that anyone ordering large quantities of precursor chemicals for chemical weapons is not going to get away with it because the corporation is required, by the chemical weapons treaty, to know its customer and to report any unusual attempts that are made to purchase large quantities of certain dangerous products.

我认为，有可能建立一种相当有效的制度，防止大量民用技术被转用于制造自主武器。坏事仍会发生，我认为这可能是不可避免的，因为自制自主武器的数量很少，制造起来可能总是可行的。然而，数量少的自主武器并不比有人驾驶的武器有太大优势。如果你要用十到二十件武器发动攻击，你不妨驾驶它们，因为你很可能能找到十到二十个人来做这件事。

I think it will be possible to have a fairly effective regime that could prevent very large diversions of civilian technology to create autonomous weapons. Bad things would still happen, and I think this may be inevitable, because in small numbers it will likely always be feasible for homemade autonomous weapons to be built. In small numbers, though, autonomous weapons don’t have a huge advantage over a piloted weapon. If you’re going to launch an attack with ten or twenty weapons, you might as well pilot them because you can probably find ten or twenty people to do that.

当然，人工智能和战争还存在其他风险，比如，当机器误解某些信号并开始相互攻击时，人工智能系统可能会意外升级战争。而未来网络渗透的风险意味着，你可能认为自己拥有基于自主武器的强大防御能力，但事实上，你的所有武器都已受到损害，当冲突开始时，它们会反过来攻击你。所有这些都加剧了战略不确定性，这可不是什么好事。

There are other risks of course with AI and warfare, such as where an AI system may accidentally escalate warfare when machines misinterpret some signal and start attacking each other. And the future risk of a cyber-infiltration means that you may think you have a robust defense based on autonomous weapons when in fact, all your weapons have been compromised and are going to turn on you instead when a conflict begins. So that all contributes to strategic uncertainty, which is not great at all.

马丁·福特：这些场景太可怕了。你还制作了一部名为《屠杀机器人》的短片，这是一部相当恐怖的视频。

MARTIN FORD: These are scary scenarios. You’ve also produced a short film called Slaughterbots, which is quite a terrifying video.

斯图尔特·J·拉塞尔：我们制作这个视频只是为了说明这些概念，因为我觉得，尽管我们尽了最大努力来描述它们并进行演示，但信息似乎并没有传达出去。人们仍然在说：“哦，自主武器是科幻小说。”他们仍然把它想象成天网和终结者，一种不存在的技术。所以，我们只是想指出，我们谈论的不是自发的邪恶武器，也不是统治世界——但我们也不再谈论科幻小说了。

STUART J. RUSSELL: We made the video really just to illustrate these concepts because I felt that, despite our best efforts to write about them and give presentations about them, that somehow the message wasn’t getting through. People were still saying, “oh, autonomous weapons are science fiction.” They were still imagining it as Skynet and Terminators, as a technology that doesn’t exist. So, we were simply trying to point out that we’re not talking about spontaneously evil weapons, and we’re not talking about taking over the world—but we also not talking about science fiction any more.

这些人工智能战争技术如今已可行，但它们也带来了一些新的极端风险。我们谈论的是可扩展的大规模杀伤性武器落入坏人之手。这些武器可能对人类造成巨大伤害。这就是自主武器。

These AI warfare technologies are feasible today, and they bring some new kinds of extreme risks. We’re talking about scalable weapons of mass destruction falling into the wrong hands. These weapons could inflict enormous damage on human populations. So, that’s autonomous weapons.

马丁·福特：2014 年，您与已故的斯蒂芬·霍金、物理学家马克斯·泰格马克和弗兰克·威尔切克共同发表了一封信，警告我们，我们对高级人工智能带来的风险没有足够重视。值得注意的是，您是信中唯一的计算机科学家。您能讲讲这封信背后的故事，以及是什么促使您写这封信的吗？（ https://www.independent.co.uk/news/science/stephen-hawking-transcendence-looks-at-the-implications-of-artificial-intelligence-but-are-we-taking-9313474.html）

MARTIN FORD: In 2014, you published a letter, along with the late Stephen Hawking and the physicists Max Tegmark and Frank Wilczek, warning that we aren’t taking the risks associated with advanced AI seriously enough. It’s notable that you were the only computer scientist among the authors. Could you tell the story behind that letter and what led you to write it? (https://www.independent.co.uk/news/science/stephen-hawking-transcendence-looks-at-the-implications-of-artificial-intelligence-but-are-we-taking-9313474.html)

斯图尔特·J·拉塞尔：这是一个有趣的故事。事情开始于我接到美国全国公共广播电台的电话，他们想采访我一部名为《超验骇客》的电影。当时我住在巴黎，这部电影还没有在巴黎上映，所以我还没看过。

STUART J. RUSSELL: So, it’s an interesting story. It started when I got a call from National Public Radio, who wanted to interview me about this movie called Transcendence. I was living in Paris at the time and the movie wasn’t out in Paris, so I hadn’t seen it yet.

我在从冰岛参加会议回来的途中恰好在波士顿停留，所以我在波士顿下了飞机，就去电影院看了这部电影。我坐在电影院前面，根本不知道电影里会发生什么，然后，“哦，看！电影里放的是伯克利计算机科学系。这有点搞笑。”约翰尼·德普扮演人工智能教授，“哦，这有点意思。”他正在做一个关于人工智能的演讲，然后有人，某个反人工智能恐怖分子决定枪杀他。所以，看到这一切发生，我不由自主地在座位上缩了缩，因为那可能真的是我当时的情况。电影的基本情节是，在他死前，他们设法将他的大脑上传到一台大型量子计算机中，这两件事的结合创造了一个超级智能实体，它威胁着要接管世界，因为它非常迅速地开发出各种令人惊叹的新技术。

I happened to have a stopover in Boston on the way back from a conference in Iceland, so I got off the plane in Boston and I went to the movie theatre to watch the movie. I’m sitting there towards the front of the theatre, and I don’t really know what’s going to happen in the movie at all and then, “Oh, look! It’s showing Berkeley computer science department. That’s kind of funny.” Johnny Depp is playing the AI professor, “Oh, that’s kind of interesting.” He’s giving a talk about AI, and then someone, some anti-AI terrorist decides to shoot him. So, I’m sort of involuntarily shrinking down in my seat seeing this happening, because that could really be me at that time. Then the basic plot of the movie is that before he dies they manage to upload his brain into a big quantum computer and the combination of those two things creates a super-intelligent entity that threatens to take over the world because it very rapidly develops all kinds of amazing new technologies.

所以，我们写了一篇文章，至少从表面上看，这是对这部电影的评论，但实际上是在说：“你知道，虽然这只是一部电影，但它所传达的信息是真实的：那就是，如果——或者当——我们创造出能够对现实世界产生主导影响的机器时，那么这会给我们带来一个非常严重的问题：事实上，我们可能会将对我们未来的控制权交给人类以外的其他实体。”

So anyway, we wrote an article that was, at least superficially, a review of the movie, but it was really saying, “You know, although this is just a movie, the underlying message is real: which is that if—or when—we create machines that can have a dominant effect on the real world, then that can present a very serious problem for us: that we could, in fact, cede control over our futures to other entities besides humans.”

问题很简单：我们的智力赋予我们控制世界的能力；因此，智力代表着对世界的权力。如果某事物的智力水平更高，那么它的力量就更大。

The problem is very straightforward: our intelligence is what gives us our ability to control the world; and so, intelligence represents power over the world. If something has a greater degree of intelligence, then it has more power.

我们已经在创造比我们强大得多的东西的道路上前进了；但无论如何，我们必须确保它们永远不会拥有任何力量。所以，当我们这样描述人工智能的情况时，人们会说，“哦，我明白了。好吧，有问题。”

We are already on the way to creating things that are much more powerful than us; but somehow, we have to make sure that they never, ever, have any power. So, when we describe the AI situation like that, people say, “Oh, I see. OK, there’s a problem.”

马丁·福特：然而，许多杰出的人工智能研究人员对这些担忧不以为然……

MARTIN FORD: And yet, a lot of prominent AI researchers are quite dismissive of these concerns...

斯图尔特·J·拉塞尔：我来谈谈这些人工智能否定论者。人们提出了各种各样的理由来说明我们为什么不应该关注人工智能问题，这些理由实在是太多了，数不胜数。我收集了大约 25 到 30 个不同的论点，但它们都有一个共同的特点，那就是它们根本就没有任何意义。它们经不起推敲。举个例子，你经常会听到这样的话：“嗯，你知道，这绝对不是问题，因为我们可以关掉它们。”这就像说在围棋比赛中击败 AlphaZero 绝对不是问题。你只要把白棋放在正确的位置，明白吗？它经不起五秒钟的推敲。

STUART J. RUSSELL: Let me talk about these AI denialists. There are various arguments that people put forward as to why we shouldn’t pay any attention to the AI problem, and that there are just too many of these arguments to count. I’ve collected somewhere between 25 and 30 distinct arguments, but they all share a single property, which is that they simply do not make any sense. They don’t really stand up to scrutiny. Just to give you one example, something you’ll often hear is, “well, you know, it’s absolutely not a problem because we’ll just be able to switch them off.” That is like saying that beating AlphaZero at Go is absolutely not a problem. You just put the white pieces in the right place, you know? It just doesn’t stand up to five seconds of scrutiny.

我认为，许多这些反对人工智能的论点都反映了一种下意识的防御反应。也许有些人会想，“我是一名人工智能研究人员。我觉得这个想法让我感到害怕，所以我要把这个想法抛在脑后，并找个理由把它抛在脑后。”这就是我关于为什么一些非常了解情况的人会试图否认人工智能会成为人类的问题的理论之一。

A lot of these AI denialist arguments, I think, reflect a kind of a knee-jerk defensive reaction. Perhaps some people think, “I’m an AI researcher. I feel threatened by this thought, and therefore I’m going to keep this thought out of my head and find some reason to keep it out of my head.” That’s one of my theories about why some otherwise very informed people will try to deny that AI is going to become a problem for humans.

这种观点甚至延伸到了人工智能社区中一些否认人工智能会成功的主流人士，这很讽刺，因为我们花了 60 年时间反击那些否认人工智能会成功的哲学家。我们也花了 60 年时间一次又一次地展示和证明，哲学家们所说的不可能的事情确实会发生——比如击败国际象棋世界冠军。

This even extends to some mainstream people in the AI community who deny that AI will ever be successful, which is ironic because we’ve spent 60 years fending off philosophers, who have denied that the AI field will ever be successful. We’ve also spent those 60 years demonstrating and proving, one time after another, how things that the philosophers said would be impossible, can indeed happen—such as beating the world champion in chess.

现在，人工智能领域的一些人突然说人工智能永远不会成功，所以没有什么可担心的。

Now, suddenly some people in the AI field are saying that AI is never going to succeed, and so there isn’t anything to worry about.

在我看来，这是一种完全病态的反应。就像核能和原子武器一样，明智的做法是假设人类的聪明才智确实会克服障碍，实现足以至少潜在地威胁放弃控制权的智能。明智的做法是做好准备，并试图找出如何设计系统来避免这种情况发生。所以这就是我的目标：帮助我们为人工智能威胁做好准备。

This is a completely pathological reaction if you ask me. It seems prudent, just as with nuclear energy and atomic weapons, to assume that human ingenuity will, in fact, overcome the obstacles and achieve intelligence of a kind that’s sufficient to present at least, potentially the threat of ceding control. It seems prudent to prepare for that and try to figure out how to design systems in such a way that that can’t happen. So that’s my goal: to help us prepare for the artificial intelligence threat.

马丁·福特：我们应该如何应对这一威胁？

MARTIN FORD: How should we address that threat?

斯图尔特·J·拉塞尔：问题的关键在于我们对人工智能的定义有点错误，因此我重新构建了人工智能的定义，如下。

STUART J. RUSSELL: The key to the problem is that we have made a slight mistake in the way that we define AI, and so I have a reconstructed a new definition for AI that goes as follows.

首先，如果我们想要制造人工智能，我们最好弄清楚智能的含义。这意味着我们必须借鉴数千年的传统、哲学、经济学和其他学科。智能的概念是，人类的智能程度取决于他们的行为是否能够实现其目标。这种概念有时被称为理性行为；它包含各种子类型的智能，如推理能力、计划能力、感知能力等等。这些都是在现实世界中智能行动所需的能力。

First of all, if we want to build artificial intelligence, we’d better figure out what it means to be intelligent. This means that we must draw from thousands of years of tradition, philosophy, economics and other disciplines. The idea of intelligence is that a human being is intelligent to the extent that their actions can be expected to achieve their objectives. This is the idea sometimes called rational behavior; and it contains within it various sub-kinds of intelligence, like the ability to reason; the ability to plan; the ability to perceive; and so on. Those are all kind of required capabilities for acting intelligently in the real world.

问题在于，如果我们成功创造出具有这些能力的人工智能和机器，那么除非它们的目标恰好与人类的目标完全一致，否则我们创造的东西就非常聪明，但其目标与我们不同。然后，​​如果人工智能比我们更聪明，那么它就会实现它的目标——而我们，可能不会！

The problem with that is that if we succeed in creating artificial intelligence and machines with those abilities, then unless their objectives happen to be perfectly aligned with those of humans, then we’ve created something that’s extremely intelligent, but with objectives that are different from ours. And then, if that AI is more intelligent than us, then it’s going to attain its objectives—and we, probably, are not!

人工智能对人类的负面影响是无穷无尽的。错误在于我们将“智能”这一对人类来说有意义的概念转移到了机器身上。

The negative consequences for humans are without limit. The mistake is in the way we have transferred the notion of intelligence, a concept that makes sense for humans, over to machines.

我们不需要拥有我们这种智能的机器。我们实际上想要的是那些能够实现我们的目标而不是它们目标的机器。

We don’t want machines with our type of intelligence. We actually want machines whose actions can be expected to achieve our objectives, not their objectives.

我们对人工智能的最初想法是，为了制造一台智能机器，我们应该构建优化器：当我们给它一个目标时，它会选择非常好的动作。然后它就会开始实现我们的目标。这可能是一个错误。到目前为止，它一直有效——但这只是因为我们还没有制造出非常智能的机器，而且我们制造的机器只放在迷你世界中，比如模拟棋盘、模拟围棋盘等等。

The original idea we had for AI was that to make an intelligent machine, we should construct optimizers: things that choose actions really well when we give them an objective. Then off it goes and achieves our objective. That’s probably a mistake. It’s worked up to now—but only because we haven’t made very intelligent machines, and the ones we have made we’ve only put in mini-worlds, like the simulated chessboard, the simulated Go board, and so on.

当人类迄今为止创造的人工智能进入现实世界时，事情就会出错，闪电崩盘就是一个例子。在闪电崩盘中，出现了一系列交易算法，其中一些相当简单，但有些则相当复杂，是基于人工智能的决策和学习系统。在现实世界中，在闪电崩盘期间，事情发生了灾难性的错误，这些机器导致股市崩盘。它们在几分钟内抹去了价值超过一万亿美元的股票。闪电崩盘是对我们的人工智能发出的警告信号。

When the AI that humans have so far made, get out into the real-world, that’s when things can go wrong, and we saw an example of this with the flash crash. With the flash crash, there was a bunch of trading algorithms, some of them fairly simple, but some of them fairly complicated AI-based decision-making and learning systems. Out there in the real world, during the flash crash things went catastrophically wrong and those machines crashed the stock market. They eliminated more than a trillion dollars of value in equities in the space of a few minutes. The flash crash was a warning signal about our AI.

思考人工智能的正确方式是，我们应该制造能够帮助我们实现目标的机器，但我们绝对不会将我们的目标直接放入机器中！

The right way to think about AI is that we should be making machines which act in ways to help us achieve our objectives through them, but where we absolutely do not put our objectives directly into the machine!

我的观点是，人工智能必须始终被设计为帮助我们实现目标，但不应假设人工智能系统知道这些目标是什么。

My vision is that AI must always be designed to try to help us achieve our objectives, but that AI systems should not be assumed to know what those objectives are.

如果我们以这种方式制造人工智能，那么人工智能必须追求的目标的性质就总是存在明显的不确定性。事实证明，这种不确定性实际上是我们所需的安全边际。

If we make AI this way, then there is always an explicit uncertainty about the nature of the objectives that an AI is obliged to pursue. It turns out that this uncertainty actually is the margin of safety that we require.

我举个例子来说明我们确实需要的安全边际。让我们回到一个古老的观念，即如果我们遇到麻烦，我们可以——如果需要的话——直接关掉机器。当然，如果机器有一个目标，比如“去拿咖啡”，那么一台足够智能的机器显然会意识到，如果有人把它关掉，它就无法去拿咖啡了。如果它的使命，如果它的目标是去拿咖啡，那么从逻辑上讲，它会采取措施防止自己被关掉。它会禁用关闭开关。它可能会压制任何试图关掉它的人。所以，当你有一台足够智能的机器时，你可以想象一个简单的目标“去拿咖啡”会带来所有这些意想不到的后果。

I’ll give you an example to demonstrate this margin of safety that we really do need. Let’s go back to an old idea that we can—if we ever need to—just switch the machine off if we get into trouble. Well, of course, you know, if the machine has an objective like, “fetch the coffee,” then obviously a sufficiently intelligent machine realizes that if someone switches it off, then it’s not going to be able to fetch the coffee. If its life’s mission, if its objective, is to fetch the coffee, then logically it will take steps to prevent itself from being switched off. It will disable the Off switch. It will possibly neutralize anyone who might attempt to switch it off. So, you can imagine all these unanticipated consequences of a simple objective like “fetch the coffee,” when you have a sufficiently intelligent machine.

现在，在我的 AI 愿景中，我们设计了这样一种机器，尽管它仍然想要“取咖啡”，但它明白人类可能关心很多其他事情，但它并不知道这些事情是什么！在这种情况下，AI 明白它可能会做一些人类不喜欢的事情——如果人类关闭它，那是为了防止一些会让人类不高兴的事情。因为在这个愿景中，机器的目标是避免让人类不高兴，即使 AI 不知道这意味着什么，它实际上也有动机让自己被关闭。

Now in my vision for AI, we instead design the machine so that although it still wants to “fetch the coffee” it understands that there are a lot of other things that human beings might care about, but it doesn’t really know what those are! In that situation, the AI understands that it might do something that the human doesn’t like—and if the human switches it off, that’s to prevent something that would make the human unhappy. Since in this vision the goal of the machine is to avoid making the human unhappy, even though the AI doesn’t know what that means, it actually has an incentive to allow itself to be switched off.

我们可以将这一人工智能愿景转化为数学，并表明安全边际（在这种情况下，指的是机器允许自己关闭的动机）与它对人类目标的不确定性直接相关。当我们消除这种不确定性时，机器开始相信它确实知道真正的目标是什么，那么安全边际就会再次消失，机器最终会阻止我们关闭它。

We can take this particular vision for AI and put it into mathematics, and show that the margin of safety (meaning, in this case, the incentive that the machine has to allow itself to be switched off) is directly related to the uncertainty it has about the human objective. As we eliminate that uncertainty, and the machine starts to believe that it knows, for sure, what the true objective really is, then that margin of safety begins to disappear again, and the machine will ultimately stop us from switching it off.

通过这种方式，我们可以证明，至少在一个简化的数学框架中，当你以这种方式设计机器时（对它们要追求的目标有明确的不确定性），那么它们就可以被证明是有益的，这意味着你可以证明有这台机器比没有这台机器更好。

In this way, we can show that, at least in a simplified mathematical framework, that when you design machines this way—with explicit uncertainty about the objective that they are to pursue—then they can be provably beneficial, meaning that you are provably better off with this machine than without.

我在这里分享的内容表明，可能存在一种构思人工智能的方式，它与我们迄今为止对人工智能的思考略有不同，有方法可以构建一个在安全性和控制方面具有更好特性的人工智能系统。

What I’ve shared here is an indication that there may be a way of conceiving of AI which is a little bit different from how we’ve been thinking about AI so far, that there are ways to build an AI system that has much better properties, in terms of safety and control.

马丁·福特：与人工智能安全和控制问题相关，很多人担心与其他国家（尤其是中国）的军备竞赛。这是我们应该认真对待、应该非常关注的事情吗？

MARTIN FORD: Related to these issues of AI safety and control, a lot of people worry about an arms race with other countries, especially China. Is that something we should take seriously, something we should be very concerned about?

斯图尔特·J·拉塞尔：尼克·博斯特罗姆和其他人提出了一种担忧，如果一个政党认为人工智能领域的战略主导地位是其国家安全和经济领导地位的关键部分，那么该政党就会被迫尽快开发人工智能系统的能力，而且不会过多担心可控性问题。

STUART J. RUSSELL: Nick Bostrom and others have raised a concern that, if a party feels that strategic dominance in AI is a critical part of their national security and economic leadership, then that party will be driven to develop the capabilities of AI systems—as fast as possible, and yes, without worrying too much about the controllability issues.

从高层次来看，这听起来似乎是一个合理的论点。另一方面，当我们生产可以在现实世界中运行的人工智能产品时，会有明显的经济动机来确保它们处于控制之下。

At a high level, that sounds like a plausible argument. On the other hand, as we produce AI products that can operate out there in the real world, there will be a clear economic incentive to make sure that they remain under control.

为了探讨这种情况，让我们来设想一下可能很快就会出现的产品：一个相当智能的个人助理，它可以跟踪你的活动、对话、人际关系等，并以专业人类助理可能帮助你的方式管理你的生活。现在，如果这样的系统不能很好地理解人类的偏好，并且以我们已经讨论过的不安全的方式行事，那么人们根本就不会购买它。如果它误解了这些事情，那么它可能会为你预订每晚 20,000 美元的酒店房间，或者它可能会取消与副总裁的会面，因为你应该去看牙医。

To explore this kind of scenario, let’s think about a product that might come along fairly soon: a reasonably intelligent personal assistant that keeps track of your activities, conversations, relationships and so on, and kind of runs your life in the way that a good professional human assistant might help you. Now, if such a system does not have a good understanding of human preferences, and acts in ways that that are unsafe in ways that we’ve already talked about, then people are simply not going to buy it. If it misunderstands these things, then it might book you into a $20,000-a-night hotel room, or it might cancel a meeting with the vice president because you’re supposed to go to the dentist.

在这种情况下，人工智能会误解你的偏好，它不会谦虚地了解你的偏好，而是认为它知道你想要什么，但事实完全是错的。我在其他论坛上引用过一个家用机器人的例子，它不明白猫的营养价值远低于猫的情感价值，所以它决定把猫煮了当晚餐。如果发生这种情况，家用机器人行业就会终结。没有人会希望家里有一个会犯这种错误的机器人。

In those kinds of situations, the AI is misunderstanding your preferences and, rather than being humble about its understanding of your preferences, it thinks that it knows what you want, and it is just being plain wrong about it. I’ve cited in other forums the example of a domestic robot that doesn’t understand that the nutritional value of a cat is a lot less than the sentimental value of a cat, and so it just decides to cook the cat for dinner. If that happened, that would be the end of the domestic robot industry. No one is going to want a robot in its house that could make that kind of mistake.

如今，生产日益智能产品的人工智能公司必须至少解决这一问题的一个版本，才能使其产品成为优秀的人工智能系统。

Today, AI companies that are producing increasingly intelligent products have to solve at least a version of this problem in order for their products to be good AI systems.

我们需要让人工智能界认识到，不可控、不安全的人工智能就不是好的人工智能。

We need to get the AI community to understand that AI that is not controllable and safe, is just not good AI.

就像一座倒塌的桥梁根本不是一座好桥梁一样，我们需要认识到，不可控制和不安全的人工智能根本不是好的人工智能。土木工程师不会到处说，“哦，是的，我设计的桥梁不会倒塌，你知道，不像其他人，他设计的桥梁会倒塌。”“桥梁”一词的含义就是它不应该倒塌。

In the same way that a bridge that falls down is simply not a good bridge, we need to recognize that AI that is not controllable and safe, is just not good AI. Civil engineers don’t go around saying, “Oh yeah, I design bridges that don’t fall down, you know, unlike the other guy, he designs bridges that fall down.” It’s just built into the meaning of the word “bridge” that it’s not supposed to fall down.

这应该成为我们定义人工智能时所要表达的含义。我们需要这样定义人工智能，即人工智能应该处于它应该为之工作的人类的控制之下，无论在哪个国家。我们需要这样定义人工智能，即它现在和将来都具有我们称之为可纠正性的特性：它可以被关闭，如果它做了我们不喜欢的事情，它也可以得到纠正。

This should be built into the meaning of what we mean when we define AI. We need to define AI in such a way that it remains under the control of the humans that it’s supposed to be working for, in any country. And we need to define AI so that it has, now and in the future, properties that we call corrigibility: that it is able to be switched off, and that it is able to be corrected if it’s doing something that we don’t like.

如果我们能够让全世界人工智能领域的每个人都明白，这些只是优秀人工智能的必要特征，那么我认为我们将在使人工智能领域的未来前景更加光明方面迈出一大步。

If we can get everyone in AI, around the world, to understand that these are just necessary characteristics of good AI, then I think we move a long way forward in making the future prospects of the field of AI much, much brighter.

让人工智能领域走向灭亡的最佳方式就是出现重大控制故障，就像切尔诺贝利和福岛核事故导致核工业走向灭亡一样。如果我们无法解决控制问题，人工智能就会走向灭亡。

There’s also no better way to kill the field of AI than to have a major control failure, just as the nuclear industry killed itself through Chernobyl and Fukushima. AI will kill itself if we fail to address the control issue.

马丁·福特：那么，总的来说，您是一个乐观主义者吗？您认为事情会顺利吗？

MARTIN FORD: So, on balance, are you an optimist? Do you think that things are going to work out?

斯图尔特·J·拉塞尔：是的，我确实认为自己是个乐观主义者。我认为还有很长的路要走。我们只是触及了这个控制问题的表面，但第一步似乎很有成效，因此我相当乐观地认为，人工智能的发展道路将引领我们走向我们所谓的“可证明有益的人工智能系统”。

STUART J. RUSSELL: Yes, I do think that I’m an optimist. I think there’s a long way to go. We are just scratching the surface of this control problem, but the first scratching seems to be productive, and so I’m reasonably optimistic that there is a path of AI development that leads us to what we might describe as “provably beneficial AI systems.”

当然，即使我们解决了控制问题，即使我们确实建立了可证明有益的人工智能系统，也存在风险，有些方面会选择不使用它们。这里的风险是，一方或另一方只选择放大人工智能的能力，而不考虑安全方面。

Of course, there is the risk that even if we do solve the control problem and even if we do build provably beneficial AI systems, that there will be some parties who choose not to use them. The risk here is that one party or another chooses only to magnify the capabilities of AI without regarding the safety aspects.

这可能是邪恶博士的角色类型，是《王牌大贱谍》中的反派，他想要统治世界，却意外地发布了一个人工智能系统，最终给所有人带来了灾难。或者这可能是一个更具社会性的风险，一开始，拥有有能力、可控的人工智能对社会来说是一件好事，但后来我们过度使用它了。在这些风险情景中，我们走向了一个衰弱的人类社会，我们把太多的知识和太多的决策权都转移到了机器上，我们永远无法恢复。在这条社会道路上，我们最终可能会失去作为人类的全部自主权。

This could be the Dr. Evil character type, the Austin Powers villain who wants to take over the world and accidentally releases an AI system that ends up being catastrophic for everyone. Or it could be a much more sociological risk, where it starts off as very nice for society to have capable, controllable AI but we then overuse it. In those risk scenarios, we head towards an enfeebled human society where we’ve moved too much of our knowledge and too much of our decision-making into machines, and we can never recover it. We could eventually lose our entire agency as humans along this societal path.

这种社会图景在电影《机器人总动员》中描绘的未来世界中，人类登上宇宙飞船，由机器照看。人类逐渐变得更胖、更懒、更愚蠢。这是科幻小说中的一个老主题，在电影《机器人总动员》中得到了非常清晰的体现。这是我们需要关注的未来，前提是我们能够成功应对我们讨论过的所有其他风险。

This societal picture is how the future is depicted in the WALL-E movie, where humanity is off on spaceships and being looked after by machines. Humanity gradually becomes fatter and lazier and stupider. That’s an old theme in science fiction and it’s very clearly illustrated in the WALL-E movie. That is a future that we need to be concerned about, assuming we successfully navigate all the other risks that we’ve been discussing.

作为一个乐观主义者，我也能预见到未来人工智能系统设计得足够好，它们会对人类说：“别利用我们。自己去学习吧。保持自己的能力，通过人类而不是机器来传播文明。”

As an optimist, I can also see a future where AI systems are well enough designed that they’re saying to humans, “Don’t use us. Get on and learn stuff yourself. Keep your own capabilities, propagate civilization through humans, not through machines.”

当然，如果我们作为一个种族被证明过于懒惰和贪婪，我们可能仍然会忽略一个有帮助且设计良好的人工智能；然后我们将付出代价。从这个意义上说，这真的可能成为一个更大的社会文化问题，我确实认为我们需要作为人类做好准备，确保这种情况不会发生。

Of course, we might still ignore a helpful and well-design AI, if we prove to be too lazy and greedy as a race; and then we’ll pay the price. In that sense, this really might become more of a sociocultural problem, and I do think that we need to do work as a human race to prepare and make sure this doesn’t happen.

斯图尔特·J·拉塞尔 是加州大学伯克利分校电气工程和计算机科学教授，被公认为人工智能领域的全球领军人物之一。他与彼得·诺维格合著了《人工智能：一种现代方法》，该书是目前在 118 个国家/地区的 1300 多所高校中使用的领先人工智能教科书。

STUART J. RUSSELL is a professor of electrical engineering and computer science at the University of California Berkeley and is widely recognized as one of the world’s leading contributors in the field of artificial intelligence. He is the co-author, along with Peter Norvig, of Artificial Intelligence: A Modern Approach, which is the leading AI textbook currently in use at over 1300 colleges and universities in 118 countries.

斯图尔特于 1982 年获得牛津大学瓦德汉学院物理学学士学位，并于 1986 年获得斯坦福大学计算机科学博士学位。他的研究涵盖了与人工智能相关的许多主题，例如机器学习、知识表示和计算机视觉，并获得过无数奖项和荣誉，包括 IJCAI 计算机与思想奖以及当选为美国科学促进会、人工智能促进会和计算机协会会员。

Stuart received his undergraduate degree in Physics from Wadham College, Oxford in 1982 and his PhD in Computer Science from Stanford in 1986. His research has covered many topics related to AI, such as machine learning, knowledge representation, and computer vision, and he has received numerous awards and distinctions, including the IJCAI Computers and Thought Award and election as a fellow to the American Association for the Advancement of Science, the Association for the Advancement of Artificial Intelligence and the Association of Computing Machinery.

多伦多大学计算机科学名誉杰出教授谷歌副总裁兼工程研究员

EMERITUS DISTINGUISHED PROFESSOR OF COMPUTER SCIENCE, UNIVERSITY OF TORONTO VICE PRESIDENT & ENGINEERING FELLOW, GOOGLE

杰弗里·辛顿有时被称为深度学习教父，他是深度学习一些关键技术的推动者，例如 反向传播、玻尔兹曼机和 Capsules 神经网络。除了在谷歌和多伦多大学任职外，他还是人工智能向量研究所的首席科学顾问。

Geoffrey Hinton is sometimes known as the Godfather of Deep Learning, and he has been the driving force behind some of its key technologies, such as backpropagation, Boltzmann machines, and the Capsules neural network. In addition to his roles at Google and the University of Toronto, he is also Chief Scientific Advisor of the Vector Institute for Artificial Intelligence.

马丁·福特：你因研究反向传播算法而出名。你能解释一下反向传播是什么吗？

MARTIN FORD: You’re most famous for working on the backpropagation algorithm. Could you explain what backpropagation is?

杰弗里·辛顿：解释神经网络的最好方式就是解释它不是什么。大多数人在思考神经网络时，都会有一个明显的训练算法：假设你有一个包含多层神经元的网络，底层有一个输入，顶层有一个输出。每个神经元与每个连接都关联一个权重。每个神经元所做的就是查看下一层中的神经元，并将下一层中神经元的活动乘以权重，然后将所有这些相加，并给出一个输出，该输出是该和的函数。通过调整连接的权重，你可以让网络做任何你想做的事情，比如查看猫的图片并将其标记为猫。

GEOFFREY HINTON: The best way to explain it is by explaining what it isn’t. When most people think about neural networks, there’s an obvious algorithm for training them: Imagine you have a network that has layers of neurons, and you have an input at the bottom layer, and an output at the top layer. Each neuron has a weight associated with each connection. What each neuron does is look at the neurons in the layer below and it multiplies the activity of a neuron in the layer below by the weight, then adds all that up and gives an output that’s a function of that sum. By adjusting the weights on the connections, you can get networks that do anything you like, such as looking at a picture of a cat and labeling it as a cat.

问题是，如何调整权重，让网络按照你的意愿运行？有一种非常简单的算法确实可行，但速度非常慢——这是一种愚蠢的变异算法——你从所有连接的随机权重开始，然后为网络提供一组示例，看看它的效果如何。然后，你取其中一个权重，稍作修改，现在再给它一组示例，看看它的效果是比以前更好还是更差。如果效果比以前更好，则保留所做的更改。如果效果比以前更差，则不要保留该更改，或者你可以将权重朝相反方向更改。然后，你取另一个权重，并执行相同的操作。

The question is, how should you adjust the weights so that the network does what you want? There’s a very simple algorithm that will actually work but is incredibly slow—it’s a dumb mutation algorithm—where you start with random weights on all the connections, and you give your network a set of examples and see how well it works. You then take one of those weights, and you change it a little bit, and now you give it another set of examples to see if it works better or worse than it did before. If it works better than it did before, you keep the change you made. If it works worse than it did before, you don’t keep that change, or perhaps you change the weight in the opposite direction. Then you take another weight, and you do the same thing.

你必须遍历所有的权重，对于每个权重，你必须测量网络在一组示例上的表现，每个权重都必须更新多次。这是一个非常慢的算法，但它有效，它会做任何你想做的事。

You have to go around all of the weights, and for each weight, you have to measure how well the network does on a set of examples, with each weight having to be updated multiple times. It is an incredibly slow algorithm, but it works, and it’ll do whatever you want.

反向传播基本上是实现相同目标的一种方法。它是一种调整权重的方法，以便网络按照你的意愿运行，但与愚蠢算法不同，它的速度要快得多。它的速度是网络中权重数量的倍数。如果你有一个包含十亿个权重的网络，反向传播将比愚蠢算法快十亿倍。

Backpropagation is basically a way of achieving the same thing. It’s a way of tinkering with the weights so that the network does what you want, but unlike the dumb algorithm, it’s much, much faster. It’s faster by a factor of how many weights there are in the network. If you’ve got a network with a billion weights, backpropagation is going to be a billion times faster than the dumb algorithm.

愚蠢算法的工作原理是让你稍微调整其中一个权重，然后测量网络的表现如何。对于进化来说，这就是你必须做的，因为从基因到成品的过程取决于你所处的环境。你无法根据基因型准确预测表现型会是什么样子，也无法预测表现型会有多成功，因为这取决于世界上发生的事情。

The dumb algorithm works by having you adjust one of the weights slightly, followed by you measuring to see how well the network does. For evolution, that’s what you’ve got to do because the process that takes you from your genes to the finished product depends on the environment you’re in. There’s no way you can predict exactly what the phenotype will look like from the genotype, or how successful the phenotype will be because that depends on what’s going on in the world.

然而，在神经网络中，处理器会从输入和权重开始，判断你是否能成功产生正确的输出。你可以控制整个过程，因为这一切都是在神经网络内部进行的；你知道所有涉及的权重。反向传播利用所有这些信息，通过网络向后发送信息。利用它知道所有权重的事实，它可以并行计算网络中的每个权重，无论你是否应该将其调大或调小以改善输出。

In a neural net, however, the processor takes you from the input and the weights to how successful you are in producing the right output. You have control over that whole process because it’s all going on inside the neural net; you know all the weights that are involved. Backpropagation makes use of all that by sending information backward through the net. Using the fact that it knows all the weights, it can compute in parallel for every single weight in the network, whether you should make it a little bit bigger or smaller to improve the output.

不同之处在于，在进化中，你要测量改变的影响，而在反向传播中，你要计算改变的影响，而且你可以同时对所有权重进行此操作，而不会受到干扰。使用反向传播，你可以快速调整权重，因为你可以给它一些例子，然后反向传播它所说的和它应该说的之间的差异，现在你可以弄清楚如何同时改变所有权重，使它们都变得更好一点。你仍然需要多次执行这个过程，但它比进化方法快得多。

The difference is that in evolution, you measure the effect of a change, and in backpropagation, you compute what the effect would be of making a change, and you can do that for all the weights at once with no interference. With backpropagation you can adjust the weights rapidly because you can give it a few examples, then backpropagate the discrepancies between what it said and what it should have said, and now you can figure out how to change all of the weights simultaneously to make all of them a little bit better. You still need to do the process a number of times, but it’s much faster than the evolutionary approach.

马丁·福特：反向传播算法最初是由大卫·鲁梅尔哈特创建的，对吗，然后您继续了这项工作？

MARTIN FORD: The backpropagation algorithm was originally created by David Rumelhart, correct, and you took that work forward?

杰弗里·辛顿：在大卫·鲁梅尔哈特之前，许多人发明了不同版本的反向传播。它们主要是独立发明，我觉得我在这方面的功劳太大了。我看到媒体说我发明了反向传播，这完全是错误的。这是学者觉得自己在某些事情上功劳太大的罕见情况之一！我的主要贡献是展示了如何使用它来学习分布式表示，所以我想澄清一下。

GEOFFREY HINTON: Lots of different people invented different versions of backpropagation before David Rumelhart. They were mainly independent inventions, and it’s something I feel I’ve got too much credit for. I’ve seen things in the press that say I invented backpropagation, and that’s completely wrong. It’s one of these rare cases when an academic feels he’s got too much credit for something! My main contribution was to show how you can use it for learning distributed representations, so I’d like to set the record straight on that.

1981 年，我在加利福尼亚州圣地亚哥做博士后，David Rumelhart 提出了反向传播的基本思想，所以这是他的发明。我和 Ronald Williams 与他一起努力，正确地制定了它。我们让它工作起来，但我们没有做出任何特别令人印象深刻的事情，也没有发表任何东西。之后，我去了卡内基梅隆大学，研究玻尔兹曼机，我认为这是一个更有趣的想法，尽管它的效果并不好。然后在 1984 年，我回去再次尝试反向传播，以便将其与玻尔兹曼机进行比较，发现它实际上效果更好，所以我又开始与 David Rumelhart 交流。

In 1981, I was a postdoc in San Diego, California and David Rumelhart came up with the basic idea of backpropagation, so it’s his invention. Myself and Ronald Williams worked with him on formulating it properly. We got it working, but we didn’t do anything particularly impressive with it, and we didn’t publish anything. After that, I went off to Carnegie Mellon and worked on the Boltzmann machine, which I think of as a much more interesting idea, even though it doesn’t work as well. Then in 1984, I went back and tried backpropagation again so I could compare it with the Boltzmann machine, and discovered it actually worked much better, so I started communicating with David Rumelhart again.

让我对反向传播真正感到兴奋的是所谓的家谱任务，在这个任务中，你可以展示反向传播可以学习分布式表示。我从高中开始就对大脑的分布式表示很感兴趣，现在我们终于有了一种学习它们的高效方法！如果你给它一个问题，比如我要输入两个单词，它必须输出与之相关的第三个单词，它就会学习这些单词的分布式表示，而这些分布式表示将捕捉这些单词的含义。

What got me really excited about backpropagation was what I called the family trees task, where you could show that backpropagation can learn distributed representations. I had been interested in the brain having distributed representations since high school, and finally, we had an efficient way to learn them! If you gave it a problem, such as if I was to input two words and it has to output the third word that goes with that, it would learn distributed representations for the words, and those distributed representations would capture the meanings of the words.

早在 20 世纪 80 年代中期，计算机还很慢，我曾举过一个简单的例子，你有一棵家谱，我会告诉你家谱中的关系。我会告诉你一些事情，比如夏洛特的母亲是维多利亚，所以我会说夏洛特和母亲，正确答案是维多利亚。我还会说夏洛特和父亲，正确答案是詹姆斯。一旦我说了这两件事，因为这是一棵非常正常的家谱，没有离婚，你就可以使用传统的人工智能，根据你对家庭关系的了解推断维多利亚一定是詹姆斯的配偶，因为维多利亚是夏洛特的母亲，詹姆斯是夏洛特的父亲。神经网络也可以推断出这一点，但它不是通过使用推理规则来做到这一点的，而是通过学习每个人的一系列特征来实现的。维多利亚和夏洛特都是一组独立的特征，然后通过使用这些特征向量之间的交互，这将导致输出成为正确人选的特征。它可以基于 Charlotte 的特征和 Mother 的特征得出 Victoria 的特征，当你训练它时，它就会学会这样做。最令人兴奋的是，对于这些不同的单词，它会学习这些特征向量，并且它正在学习单词的分布式表示。

Back in the mid-1980s, when computers were very slow, I used a simple example where you would have a family tree, and I would tell you about relationships within that family tree. I would tell you things like Charlotte’s mother is Victoria, so I would say Charlotte and mother, and the correct answer is Victoria. I would also say Charlotte and father, and the correct answer is James. Once I’ve said those two things, because it’s a very regular family tree with no divorces, you could use conventional AI to infer using your knowledge of family relations that Victoria must be the spouse of James because Victoria is Charlotte’s mother and James is Charlotte’s father. The neural net could infer that too, but it didn’t do it by using rules of inference, it did it by learning a bunch of features for each person. Victoria and Charlotte would both be a bunch of separate features, and then by using interactions between those vectors of features, that would cause the output to be the features for the correct person. From the features for Charlotte and from the features for mother, it could derive the features for Victoria, and when you trained it, it would learn to do that. The most exciting thing was that for these different words, it would learn these feature vectors, and it was learning distributed representations of words.

1986 年，我们向《自然》杂志提交了一篇论文，其中有反向传播学习单词分布式特征的例子，我与这篇论文的一位审稿人进行了交谈，他对此感到非常兴奋，因为这个系统正在学习这些分布式表征。他是一名心理学家，他明白，拥有一个可以学习事物表征的学习算法是一个重大突破。我的贡献不是发现反向传播算法，那是鲁梅尔哈特已经弄清楚的事情，而是表明反向传播可以学习这些分布式表征，这正是心理学家感兴趣的地方，最终也让人工智能专家感兴趣。

We submitted a paper to Nature in 1986 that had this example of backpropagation learning distributed features of words, and I talked to one of the referees of the paper, and that was what got him really excited about it, that this system was learning these distributed representations. He was a psychologist, and he understood that having a learning algorithm that could learn representations of things was a big breakthrough. My contribution was not discovering the backpropagation algorithm, that was something Rumelhart had pretty much figured out, it was showing that backpropagation would learn these distributed representations, and that was what was interesting to psychologists, and eventually, to AI people.

几年后，在 20 世纪 90 年代初，Yoshua Bengio 重新发现了同一种网络，但当时计算机的速度更快。Yoshua 将其应用于语言，因此他会获取真实文本，以几个单词作为上下文，然后尝试预测下一个单词。他展示了神经网络在这方面表现得相当好，它可以发现单词的这些分布式表示。它产生了巨大的影响，因为反向传播算法可以学习表示，而你不必手动输入它们。Yann LeCun 等人已经在计算机视觉领域这样做了一段时间。他展示了反向传播将学习用于处理视觉输入的良好过滤器，以便做出正确的决策，这更明显一些，因为我们知道大脑会做这样的事情。反向传播将学习捕捉单词含义和语法的分布式表示，这一事实是一个重大突破。

Quite a few years later, in the early 1990s, Yoshua Bengio rediscovered the same kind of network but at a time where computers were faster. Yoshua was applying it to language, so he would take real text, taking a few words as context, and then try and predict the next word. He showed that the neural network was pretty good at that and that it would discover these distributed representations of words. It made a big impact because the backpropagation algorithm could learn representations and you didn’t have to put them in by hand. People like Yann LeCun had been doing that in computer vision for a while. He was showing that backpropagation would learn good filters for processing visual input in order to make good decisions, and that was a bit more obvious because we knew the brain did things like that. The fact that backpropagation would learn distributed representations that captured the meanings and the syntax of words was a big breakthrough.

马丁·福特：当时，神经网络的使用还不是人工智能研究的重点，这种说法对吗？直到最近，这一领域才开始受到重视。

MARTIN FORD: Is it correct to say that at that time using neural networks was still not a primary thrust of AI research? It’s only quite recently this has come to the forefront.

GEOFFREY HINTON：这话有一定道理，但你也需要区分人工智能和机器学习与心理学。1986 年反向传播流行起来后，很多心理学家开始对它感兴趣，他们并没有真正失去兴趣，他们一直认为这是一种有趣的算法，也许不是大脑所做的，而是一种开发表征的有趣方式。偶尔，你会看到只有少数人在研究它，但事实并非如此。在心理学领域，很多人对它感兴趣。人工智能领域的情况是，在 20 世纪 80 年代末，Yann LeCun 在识别手写数字方面取得了令人印象深刻的成果，反向传播还有其他各种令人印象深刻的应用，从语音识别到预测信用卡欺诈。然而，反向传播的支持者认为它会带来惊人的效果，他们可能确实夸大了它。它确实没有达到我们的期望。我们原以为它会很棒，但实际上，它只是相当不错。

GEOFFREY HINTON: There’s some truth to that, but you also need to make a distinction between AI and machine learning on the one hand, and psychology on the other hand. Once backpropagation became popular in 1986, a lot of psychologists got interested in it, and they didn’t really lose their interest in it, they kept believing that it was an interesting algorithm, maybe not what the brain did, but an interesting way of developing representations. Occasionally, you see the idea that there were only a few people working on it, but that’s not true. In psychology, lots of people stayed interested in it. What happened in AI was that in the late 1980s, Yann LeCun got something impressive working for recognizing handwritten digits, and there were various other moderately impressive applications of backpropagation from things like speech recognition to predicting credit card fraud. However, the proponents of backpropagation thought it was going to do amazing things, and they probably did oversell it. It didn’t really live up to the expectations we had for it. We thought it was going to be amazing, but actually, it was just pretty good.

在 20 世纪 90 年代初期，其他机器学习方法在小型数据集上的表现优于反向传播，而且只需要进行更少的调整就能发挥良好作用。特别是，一种称为支持向量机的算法在识别手写数字方面比反向传播更好，而手写数字一直是反向传播效果非常好的经典例子。正因为如此，机器学习社区对反向传播失去了兴趣。他们认为反向传播涉及太多调整，效果不值得如此调整，而且认为仅从输入和输出中就能学习多层隐藏表示是毫无希望的。每一层都是一大堆以特定方式表示的特征检测器。

In the early 1990s, other machine learning methods on small datasets turned out to work better than backpropagation and required fewer things to be fiddled with to get them to work well. In particular, something called the support vector machine did better at recognizing handwritten digits than backpropagation, and handwritten digits had been a classic example of backpropagation doing something really well. Because of that, the machine learning community really lost interest in backpropagation. They decided that there was too much fiddling involved, it didn’t work well enough to be worth all that fiddling, and it was hopeless to think that just from the inputs and outputs you could learn multiple layers of hidden representations. Each layer would be a whole bunch of feature detectors that represent in a particular way.

反向传播的理念是，你可以学习很多层，然后你就能做出令人惊叹的事情，但我们很难学习超过几层，而且我们无法做出令人惊叹的事情。统计学家和人工智能界人士的普遍共识是，我们只是一厢情愿。我们认为，仅从输入和输出，你就应该能够学习所有这些权重；但这是不现实的。你必须掌握大量知识才能让任何事情奏效。

The idea of backpropagation was that you’d learn lots of layers, and then you’d be able to do amazing things, but we had great difficulty learning more than a few layers, and we couldn’t do amazing things. The general consensus among statisticians and people in AI was that we were wishful thinkers. We thought that just from the inputs and outputs, you should be able to learn all these weights; and that was just unrealistic. You were going to have to wire in lots of knowledge to make anything work.

直到 2012 年，计算机视觉领域的人都持这种观点。大多数计算机视觉领域的人都认为这种东西太疯狂了，尽管 Yann LeCun 有时会让系统比最好的计算机视觉系统运行得更好，但他们仍然认为这种东西太疯狂了，这不是做视觉的正确方法。他们甚至拒绝了 Yann 的论文，尽管他们在特定问题上的表现比最好的计算机视觉系统更好，因为审稿人认为这是错误的做法。这是科学家们说的一个很好的例子，“我们已经决定了答案应该是什么样的，任何看起来不像我们所相信的答案的东西都是没有意义的。”

That was the view of people in computer vision until 2012. Most people in computer vision thought this stuff was crazy, even though Yann LeCun sometimes got systems working better than the best computer vision systems, they still thought this stuff was crazy, it wasn’t the right way to do vision. They even rejected papers by Yann, even though they worked better than the best computer vision systems on particular problems, because the referees thought it was the wrong way to do things. That’s a lovely example of scientists saying, “We’ve already decided what the answer has to look like, and anything that doesn’t look like the answer we believe in is of no interest.”

最终，科学胜出，我的两个学生赢得了一场大型公开竞赛，而且他们赢得非常出色。他们的错误率几乎是最佳计算机视觉系统的一半，他们主要使用 Yann LeCun 实验室开发的技术，但也混合了一些我们自己的技术。

In the end, science won out, and two of my students won a big public competition, and they won it dramatically. They got almost half the error rate of the best computer vision systems, and they were using mainly techniques developed in Yann LeCun’s lab but mixed in with a few of our own techniques as well.

马丁·福特：这是 ImageNet 竞赛吗？

MARTIN FORD: This was the ImageNet competition?

GEOFFREY HINTON：是的，当时发生的事情也是科学界应该发生的事情。人们过去认为完全是胡说八道的方法现在比他们所相信的方法效果好得多，两年之内，他们都改用了这种方法。所以，对于像物体分类这样的问题，现在没有人会想到不使用神经网络就可以做到这一点。

GEOFFREY HINTON: Yes, and what happened then was what should happen in science. One method that people used to think of as complete nonsense had now worked much better than the method they believed in, and within two years, they all switched. So, for things like object classification, nobody would dream of trying to do it without using a neural network now.

马丁·福特：我记得那是 2012 年。那是深度学习的转折点吗？

MARTIN FORD: This was back in 2012, I believe. Was that the inflection point for deep learning?

GEOFFREY HINTON：对于计算机视觉来说，这是一个转折点。而对于语音识别来说，转折点则早于此。2009 年，多伦多大学的两名研究生展示了使用深度学习可以制作出更好的语音识别器。他们去了 IBM 和微软实习，第三名学生带着他们的系统去了谷歌。他们构建的基本系统得到了进一步开发，在接下来的几年里，所有这些公司的实验室都转而使用神经网络进行语音识别。最初，他们只是将神经网络用作系统的前端，但最终，他们在整个系统中都使用了神经网络。语音识别领域的许多顶尖人才在 2012 年之前就转而相信神经网络，但公众的巨大影响发生在 2012 年，当时视觉社区几乎在一夜之间发生了翻天覆地的变化，这种疯狂的方法最终取得了胜利。

GEOFFREY HINTON: For computer vision, that was the inflection point. For speech, the inflection point was a few years earlier. Two different graduate students at Toronto showed in 2009 that you could make a better speech recognizer using deep learning. They went as interns to IBM and Microsoft, and a third student took their system to Google. The basic system that they had built was developed further, and over the next few years, all these companies’ labs converted to doing speech recognition using neural nets. Initially, it was just using neural networks for the frontend of their system, but eventually, it was using neural nets for the whole system. Many of the best people in speech recognition had switched to believing in neural networks before 2012, but the big public impact was in 2012, when the vision community, almost overnight, got turned on its head and this crazy approach turned out to win.

马丁·福特：如果你现在阅读媒体报道，你会觉得神经网络和深度学习等同于人工智能——它是整个领域。

MARTIN FORD: If you read the press now, you get the impression that neural networks and deep learning are equivalent to artificial intelligence—that it’s the whole field.

杰弗里·辛顿：在我的大部分职业生涯中，我都在研究人工智能，即基于逻辑的智能系统，通过制定规则让它们能够处理符号串。人们相信这就是智能，这就是他们创造人工智能的方式。他们认为智能就是根据规则处理符号串，他们只需要弄清楚符号串是什么，规则是什么，这就是人工智能。然后还有另一种根本不是人工智能的东西，那就是神经网络。它是一种通过模仿大脑学习方式来创造智能的尝试。

GEOFFREY HINTON: For most of my career, there was artificial intelligence, which meant the logic-based idea of making intelligent systems by putting in rules that allowed them to process symbol strings. People believed that’s what intelligence was, and that’s how they were going to make artificial intelligence. They thought intelligence consists of processing symbol strings according to rules, they just had to figure out what the symbol strings were and what the rules were, and that was AI. Then there was this other thing that wasn’t AI at all, and that was neural networks. It was an attempt to make intelligence by mimicking how the brain learns.

请注意，标准 AI 对学习并不特别感兴趣。在 20 世纪 70 年代，他们总是说学习不是重点。你必须弄清楚规则是什么，以及它们操纵的符号表达式是什么，我们可以稍后再担心学习。为什么？因为重点是推理。在你弄清楚它是如何推理之前，思考学习是没有意义的。基于逻辑的人对符号推理感兴趣，而基于神经网络的人对学习、感知和运动控制感兴趣。他们试图解决不同的问题，我们认为推理是人类进化过程中很晚才出现的东西，它不是理解大脑工作原理的基本方法。它建立在为其他东西设计的东西之上。

Notice that standard AI wasn’t particularly interested in learning. In the 1970s, they would always say that learning’s not the point. You have to figure out what the rules are and what the symbolic expressions they’re manipulating are, and we can worry about learning later. Why? Because the main point is reasoning. Until you’ve figured out how it does reasoning, there’s no point thinking about learning. The logic-based people were interested in symbolic reasoning, whereas the neural network-based people were interested in learning, perception, and motor control. They’re trying to solve different problems, and we believe that reasoning is something that evolutionarily comes very late in people, and it’s not the way to understand the basics of how the brain works. It’s built on top of something that’s designed for something else.

现在的情况是，行业和政府使用“AI”来表示深度学习，因此你会发现一些非常矛盾的事情。在多伦多，我们从行业和政府那里获得了大量资金，用于建立 Vector Institute，该研究所不仅进行深度学习的基础研究，还帮助行业更好地进行深度学习，并教育人们进行深度学习。当然，其他人也想得到其中的一部分资金，另一所大学声称他们从事 AI 的人比多伦多多，并提供了引用数据作为证据。那是因为他们使用的是经典 AI。他们引用了传统 AI 的引用，表示他们应该从中获得一部分资金用于深度学习，因此，AI 含义的这种混淆非常严重。如果我们不使用“AI”一词，情况会好得多。

What’s happened now is that industry and government use “AI” to mean deep learning, and so you get some really paradoxical things. In Toronto, we’ve received a lot of money from the industry and government for setting up the Vector Institute, which does basic research into deep learning, but also helps the industry do deep learning better and educates people in deep learning. Of course, other people would like some of this money, and another university claimed they had more people doing AI than in Toronto and produced citation figures as evidence. That’s because they used classical AI. They used citations of conventional AI to say they should get some of this money for deep learning, and so this confusion in the meaning of AI is quite serious. It would be much better if we just didn’t use the term “AI.”

马丁·福特：你真的认为人工智能应该只关注神经网络，其他一切都不相关吗？

MARTIN FORD: Do you really think that AI should just be focused on neural networks and that everything else is irrelevant?

杰弗里·辛顿：我认为我们应该说，人工智能的总体思路是制造非生物的智能系统，它们是人工的，可以做一些聪明的事情。然后是人工智能在很长一段时间内的含义，有时被称为老式人工智能：使用符号表达来表示事物。对于大多数学者——至少是老一辈的学者——来说，这就是人工智能的含义：致力于操纵符号表达，以此来实现智能。

GEOFFREY HINTON: I think we should say that the general idea of AI is making intelligent systems that aren’t biological, they are artificial, and they can do clever things. Then there’s what AI came to mean over a long period, which is what’s sometimes called good old-fashioned AI: representing things using symbolic expressions. For most academics—at least, the older academics—that’s what AI means: that commitment to manipulating symbolic expressions as a way to achieve intelligence.

我认为这种老式的人工智能观念是错误的。我认为他们犯了一个非常幼稚的错误。他们认为，如果有符号输入，也有符号输出，那么符号之间一定是符号。符号之间的内容不是符号串，而是神经活动的大向量。我认为传统人工智能的基本前提是错误的。

I think that old-fashioned notion of AI is just wrong. I think they’re making a very naive mistake. They believe that if you have symbols coming in and you have symbols coming out, then it must be symbols in-between all the way. What’s in-between is nothing like strings of symbols, it’s big vectors of neural activity. I think the basic premise of conventional AI is just wrong.

马丁·福特：2017 年底，您在一次采访中表示，您对反向传播算法表示怀疑，认为应该抛弃它，我们需要从头开始。（ https://www.axios.com/artificial-intelligence-pioneer-says-we-need-to-start-over-1513305524-f619efbd-9db0-4947-a9b2-7a4c310a28fe.html）这引起了很多骚动，所以我想问您这话是什么意思。

MARTIN FORD: You gave an interview toward the end of 2017 where you said that you were suspicious of the backpropagation algorithm and that it needed to be thrown out and we needed to start from scratch. (https://www.axios.com/artificial-intelligence-pioneer-says-we-need-to-start-over-1513305524-f619efbd-9db0-4947-a9b2-7a4c310a28fe.html) That created a lot of disturbance, so I wanted to ask what you meant by that.

Geoffrey Hinton：问题在于对话的背景没有得到正确报道。我当时谈论的是试图理解大脑，我提出的问题是反向传播可能不是理解大脑的正确方法。我们并不确定，但现在有一些理由相信大脑可能不会使用反向传播。我说的是，如果大脑不使用反向传播，那么大脑正在使用的任何东西都会成为人工系统的有趣候选者。我的意思根本不是说我们应该抛弃反向传播。反向传播是所有有效的深度学习的支柱，我认为我们不应该抛弃它。

Geoffrey Hinton: The problem was that the context of the conversation wasn’t properly reported. I was talking about trying to understand the brain, and I was raising the issue that backpropagation may not be the right way to understand the brain. We don’t know for sure, but there are some reasons now for believing that the brain might not use backpropagation. I said that if the brain doesn’t use backpropagation, then whatever the brain is using would be an interesting candidate for artificial systems. I didn’t at all mean that we should throw out backpropagation. Backpropagation is the mainstay of all the deep learning that works, and I don’t think we should get rid of it.

马丁·福特：据推测，它可以进一步完善？

MARTIN FORD: Presumably, it could be refined going forward?

杰弗里·辛顿：有各种各样的方法可以改进它，也可能有其他非反向传播的算法也可以工作，但我认为我们不应该停止反向传播。那太疯狂了。

GEOFFREY HINTON: There’s going to be all sorts of ways of improving it, and there may well be other algorithms that are not backpropagation that also work, but I don’t think we should stop doing backpropagation. That would be crazy.

马丁·福特：您是如何对人工智能产生兴趣的？是什么让您开始关注神经网络？

MARTIN FORD: How did you become interested in artificial intelligence? What was the path that took you to your focus on neural networks?

杰弗里·辛顿：我的故事开始于高中，当时我有一个朋友叫英曼·哈维，他是一位非常优秀的数学家，他对大脑可能像全息图一样工作的想法很感兴趣。

GEOFFREY HINTON: My story begins at high school, where I had a friend called Inman Harvey who was a very good mathematician who got interested in the idea that the brain might work like a hologram.

马丁·福特：全息图是三维表现吗？

MARTIN FORD: A hologram being a three-dimensional representation?

杰弗里·辛顿：好吧，关于全息图的重要一点是，如果你把全息图切成两半，你得到的不是半幅图片，而是整个场景的模糊图片。在全息图中，场景信息分布在整个全息图中，这与我们习惯的非常不同。它与照片非常不同，如果你剪掉照片的一部分，你会丢失照片中该部分的信息，这不仅会使整张照片变得模糊。

GEOFFREY HINTON: Well, the important thing about a proper hologram is that if you take a hologram and you cut it in half, you do not get half the picture, but instead you get a fuzzy picture of the whole scene. In a hologram, information about the scene is distributed across the whole hologram, which is very different from what we’re used to. It’s very different from a photograph, where if you cut out a piece of a photograph you lose the information about what was in that piece of the photograph, it doesn’t just make the whole photograph go fuzzier.

英曼对人类记忆可能以这种方式运作的想法很感兴趣，即单个神经元不负责存储单个记忆。他认为，大脑中发生的事情是，你调整整个大脑中神经元之间的连接强度来存储每个记忆，这基本上是一种分布式表示。当时，全息图是分布式表示的一个明显例子。

Inman was interested in the idea that human memory might work like that, where an individual neuron is not responsible for storing an individual memory. He suggested that what’s happening in the brain is that you adjust the connection strengths between neurons across the whole brain to store each memory, and that it’s basically a distributed representation. At that time, holograms were an obvious example of distributed representation.

人们误解了分布式表示的含义，但我认为它的意思是你试图表示一些东西——可能是概念——每个概念都由一大堆神经元的活动来表示，每个神经元都参与许多不同概念的表示。这与神经元和概念之间的一对一映射非常不同。这是我第一次对大脑感兴趣。我们还对大脑如何通过调整连接强度来学习事物感兴趣，所以我基本上一直对此很感兴趣。

People misunderstand what’s meant by distributed representation, but what I think it means is you’re trying to represent some things—maybe concepts—and each concept is represented by activity in a whole bunch of neurons, and each neuron is involved in the representations of many different concepts. It’s very different from a one-to-one mapping between neurons and concepts. That was the first thing that got me interested in the brain. We were also interested in how brains might learn things by adjusting connection strengths, and so I’ve been interested in that basically the whole time.

马丁·福特：你上高中的时候？哇。那么你上大学后，你的思维是如何发展的呢？

MARTIN FORD: When you were at high school? Wow. So how did your thinking develop when you went to university?

杰弗里·辛顿：我在大学学习的科目之一是生理学。我对生理学很感兴趣，因为我想了解大脑是如何工作的。在课程快结束时，他们告诉我们神经元如何发送动作电位。有人对巨型鱿鱼轴突进行了实验，弄清楚了动作电位如何沿着轴突传播，结果发现这就是大脑的工作方式。然而，令人失望的是，他们没有任何关于事物如何表现或学习的计算模型。

GEOFFREY HINTON: One of the things I studied at university was physiology. I was excited by physiology because I wanted to learn how the brain worked. Toward the end of the course they told us how neurons send action potentials. There were experiments done on the giant squid axon, figuring out how an action potential propagated along an axon, and it turned out that was how the brain worked. It was rather disappointing to discover, however, that they didn’t have any kind of computational model of how things were represented or learned.

之后，我转而学习心理学，以为他们会告诉我大脑是如何运作的，但那是在剑桥，当时它还在从行为主义中恢复过来，所以心理学主要是关于盒子里的老鼠。当时也有一些认知心理学，但它们相当不计算，我并没有真正意识到他们会弄清楚大脑是如何运作的。

After that, I switched to psychology, thinking they would tell me how the brain worked, but this was at Cambridge, and at that time it was still recovering from behaviorism, so psychology was largely about rats in boxes. There was some cognitive psychology then but they were fairly non-computational, and I didn’t really get much sense that they were ever going to figure out how the brain worked.

在心理学课程中，我做了一个关于儿童发展的项目。我观察了两到五岁的孩子，以及他们关注不同感知属性的方式是如何随着他们的成长而变化的。这个想法是，当他们很小的时候，他们主要对颜色和纹理感兴趣，但随着年龄的增长，他们对形状的兴趣会越来越大。我做了一个实验，我会向孩子们展示三个物体，其中一个是奇怪的，例如，两个黄色圆圈和一个红色圆圈。我训练孩子们指出奇怪的那个，即使是很小的孩子也能学会这样做。

During the psychology course, I did a project on child development. I was looking at children between the ages of two and five, and how the way that they attend to different perceptual properties changes as they develop. The idea is that when they’re very young, they’re mainly interested in color and texture, but as they get older, they become more interested in shape. I conducted an experiment where I would show the children three objects, of which one was the odd one out, for example, two yellow circles and a red circle. I trained the children to point at the odd one out, something that even very young children can learn to do.

我还会用两个黄色三角形和一个黄色圆圈训练他们，然后他们必须指出圆圈，因为从形状上看，圆圈是异类。在他们用简单的例子训练过后，如果发现圆圈明显是异类，我就会给他们一个测试样本，比如黄色三角形、黄色圆圈和红色圆圈。这样做的目的是，如果他们对颜色比形状更感兴趣，那么异类就是红色圆圈，但如果他们对形状比颜色更感兴趣，那么异类就是黄色三角形。这一切都很好，对于几个孩子来说，他们要么指出形状不同的黄色三角形，要么指出颜色不同的红色圆圈。不过，我记得，当我第一次和一个聪明的五岁孩子做测试时，他指着红色圆圈说：“你把那个涂错了颜色。”

I’d also train them on two yellow triangles and one yellow circle, and then they’d have to point at the circle because that was the odd one out on shape. Once they’d been trained on simple examples where there was a clear odd one out, I would then give them a test example like a yellow triangle, a yellow circle, and a red circle. The idea was that if they were more interested in color than shape, then the odd one out would be the red circle, but if they were more interested in shape than color, then the odd one out would be the yellow triangle. That was all well and good, and for a couple of children, they pointed out either the yellow triangle that was a different shape or the red circle that was a different color. I remember, though, that when I first did the test with one bright five-year-old, he pointed at the red circle, and he said, “You’ve painted that one the wrong color.”

我试图证实的模型是一个非常愚蠢、模糊的模型，它说：“当孩子还小的时候，他们会更多地关注颜色，而随着年龄的增长，他们会更多地关注形状。”这是一个非常原始的模型，它没有说明任何事情是如何运作的，只是强调从颜色到形状的轻微变化。然后，我遇到了一个孩子，他看着他们说：“你把那个涂错了颜色。”这是一个信息处理系统，它从训练示例中学习了任务是什么，因为他认为应该有一个奇怪的，他意识到没有一个奇怪的，我一定是犯了一个错误，这个错误可能是我把那个涂错了颜色。

The model that I was trying to corroborate was a very dumb, vague model that said, “when they’re little, children attend more to color and as they get bigger, they attend more to shape.” It’s an incredibly primitive model that doesn’t say how anything works, it’s just a slight change in emphasis from color to shape. Then, I was confronted by this kid who looks at them and says, “You’ve painted that one the wrong color.” Here’s an information processing system that has learned what the task is from the training examples, and because he thinks there should be an odd one out, he realizes there isn’t a single odd one out, and that I must have made a mistake, and the mistake was probably that I painted that one the wrong color.

我测试的儿童模型根本无法达到这种复杂程度。这比心理学中的任何模型都要复杂得多。这是一个聪明的信息处理系统，可以弄清楚发生了什么，对我来说，这就是心理学的终结。与他们所处理的复杂性相比，他们所拥有的模型根本不够用。

Nothing in the model of children that I was testing allowed for that level of complexity at all. This was hugely more complex than any of the models in psychology. It was an information processing system that was smart and could figure out what was going on, and for me, that was the end of psychology. The models they had were hopelessly inadequate compared with the complexity of what they were dealing with.

马丁·福特：离开心理学领域后，您是如何进入人工智能领域的？

MARTIN FORD: After leaving the field of psychology, how did you end up going into artificial intelligence?

杰弗里·辛顿：在我进入人工智能领域之前，我是一名木匠，虽然我很喜欢这份工作，但我并不是这方面的专家。在那段时间里，我遇到了一位非常优秀的木匠，这让我非常沮丧，因此我回到了学术界。

GEOFFREY HINTON: Well, before I moved into the world of AI, I became a carpenter, and whilst I enjoyed it, I wasn’t an expert at it. During that time, I met a really good carpenter, and it was highly depressing, so because of that I went back to academia.

马丁·福特：好吧，考虑到你选择的另一条道路，你不是一位出色的木匠也许是件好事！

MARTIN FORD: Well, given the other path that opened up for you, it’s probably a good thing that you weren’t a great carpenter!

杰弗里·辛顿：在尝试做木工之后，我担任了一个心理学项目的研究助理，试图了解语言在幼儿时期是如何发展的，以及社会阶层对语言发展的影响。我负责设计一份问卷，以评估母亲对孩子语言发展的态度。我骑车来到布里斯托尔一个非常贫穷的郊区，敲开了第一位我要谈话的母亲的门。她邀请我进去，给我倒了一杯茶，然后我问了她第一个问题，那就是：“你对孩子使用语言的态度是什么？”她回答说：“如果他使用语言，我们就会打他。”这就是我作为社会心理学家的职业生涯。

GEOFFREY HINTON: Following my attempt at carpentry, I worked as a research assistant on a psychology project trying to understand how language develops in very young children, and how it is influenced by social class. I was responsible for creating a questionnaire that would assess the attitude of the mother toward their child’s language development. I cycled out to a very poor suburb of Bristol, and I knocked on the door of the first mother I was due to talk to. She invited me in and gave me a cup of tea, and then I asked her my first question, which was: “What’s your attitude towards your child’s use of language?” She replied, “If he uses language, we hit him.” So that was pretty much it for my career as a social psychologist.

之后我进入了人工智能领域，成为爱丁堡大学的人工智能研究生。我的导师是一位非常杰出的科学家，名叫克里斯托弗·隆格特-希金斯 (Christopher Longuet-Higgins)，他最初是剑桥大学的化学教授，后来转行从事人工智能研究。他对大脑的工作原理非常感兴趣，尤其是研究全息图之类的东西。他意识到计算机建模是了解大脑的方法，他正在研究这个，这就是我最初和他签约的原因。不幸的是，在我和他签约的同时，他改变了主意。他认为这些神经模型不是理解智能的方法，理解智能的真正方法是尝试理解语言。

After that I went into AI and became a graduate student in artificial intelligence at The University of Edinburgh. My adviser was a very distinguished scientist called Christopher Longuet-Higgins who’d initially been a professor of chemistry at Cambridge and had then switched fields to artificial intelligence. He was very interested in how the brain might work—and in particular, studying things like holograms. He had realized that computer modeling was the way to understand the brain, and he was working on that, and that’s why I originally signed up with him. Unfortunately for me, about the same time that I signed up with him, he changed his mind. He decided that these neural models were not the way to understand intelligence, and the actual way to understand intelligence was to try and understand language.

值得记住的是，当时有一些令人印象深刻的模型——使用符号处理——可以讨论块的排列。一位名叫特里·维诺格拉德 (Terry Winograd) 的美国计算机科学教授写了一篇非常好的论文，展示了如何让计算机理解某种语言并回答问题，并且它实际上会遵循命令。你可以对它说，“把蓝色盒子里的积木放在红色立方体的顶部”，它会理解并这样做。这只是一个模拟，但它会理解这句话。这给克里斯托弗·隆格特-希金斯 (Christopher Longuet-Higgins) 留下了深刻的印象，他希望我继续研究这个，但我想继续研究神经网络。

It’s worth remembering that at the time, there were some impressive models—using symbol processing—of systems that could talk about arrangements of blocks. An American professor of computer science called Terry Winograd wrote a very nice thesis that showed how you could get a computer to understand some language and to answer questions, and it would actually follow commands. You could say to it, “put the block that’s in the blue box on top of the red cube,” and it would understand and do that. It was only in a simulation, but it would understand the sentence. That impressed Christopher Longuet-Higgins a lot, and he wanted me to work on that, but I wanted to keep working on neural networks.

克里斯托弗是个非常正直的人，但我们对我应该做什么意见完全不同。我一直拒绝按照他说的做，但他还是让我留下来。我继续研究神经网络，最终，我写了一篇关于神经网络的论文，尽管当时神经网络的效果并不好，而且大家一致认为它们只是胡说八道。

Now, Christopher was a very honorable guy, but we completely disagreed on what I should do. I kept refusing to do what he said, but he kept me on anyway. I continued my work on neural networks, and eventually, I did a thesis on neural networks, though at the time, neural networks didn’t work very well and there was a consensus that they were just nonsense.

马丁·福特：这与马文·明斯基和西摩·派普特的《感知器》一书有什么关系？

MARTIN FORD: When was this in relation to Marvin Minsky and Seymour Papert’s Perceptrons book?

杰弗里·辛顿：那是在 70 年代初期，而明斯基和帕普特的书是在 60 年代末出版的。人工智能领域几乎所有人都认为神经网络的末日已经来临。他们认为，通过研究神经网络来理解智能就像通过研究晶体管来理解智能一样，这不是正确的做法。他们认为智能完全是由程序组成的，你必须了解大脑正在使用什么程序。

GEOFFREY HINTON: This was in the early ‘70s, and Minsky and Papert’s book came out in the late ‘60s. Almost everybody in artificial intelligence thought that was the end of neural networks. They thought that trying to understand intelligence by studying neural networks was like trying to understand intelligence by studying transistors; it just wasn’t the way to do it. They thought intelligence was all about programs, and you had to understand what programs the brain was using.

这两种范式完全不同，它们试图解决不同的问题，使用完全不同的方法和不同类型的数学。当时，根本不清楚哪种范式会成为胜利的范式。今天，有些人仍然不清楚。

These two paradigms were completely different, they aimed to try and solve different problems, and they used completely different methods and different kinds of mathematics. Back then, it wasn’t at all clear which was going to be the winning paradigm. It’s still not clear to some people today.

有趣的是，一些与逻辑最相关的人实际上相信神经网络范式。最大的例子是约翰·冯·诺依曼和艾伦·图灵，他们都认为模拟神经元的大型网络是研究智能和弄清这些事物如何运作的好方法。然而，人工智能的主导方法是受逻辑启发的符号处理。在逻辑中，你获取符号串并对其进行修改以得到新的符号串，人们认为这一定是推理的工作原理。

What was interesting, was that some of the people most associated with logic actually believed in the neural net paradigm. The biggest examples are John von Neumann and Alan Turing, who both thought that big networks of simulated neurons were a good way to study intelligence and figure out how those things work. However, the dominant approach in AI was symbol processing inspired by logic. In logic, you take symbol strings and alter them to arrive at new symbol strings, and people thought that must be how reasoning works.

他们认为神经网络太过底层，它们只与实现有关，就像晶体管是计算机中的实现层一样。他们认为你无法通过观察大脑的实现方式来理解智能，他们认为你只能通过观察智能本身来理解智能，而这正是传统的人工智能方法。

They thought neural nets were far too low-level, and that they were all about implementation, just like how transistors are the implementation layer in a computer. They didn’t think you could understand intelligence by looking at how the brain is implemented, they thought you could only understand it by looking at intelligence in itself, and that’s what the conventional AI approach was.

我认为这是一个灾难性的错误，我们现在看到的就是这种情况。深度学习的成功表明，神经网络范式实际上比基于逻辑的范式更为成功，但在 20 世纪 70 年代，人们并不这么认为。

I think it was disastrously wrong, something that we’re now seeing. The success of deep learning is showing that the neural net paradigm is actually far more successful than the logic-based paradigm, but back then in the 1970s, that was not what people thought.

马丁·福特：我看到很多媒体文章都说深度学习被过度炒作了，这种炒作可能会导致失望，进而减少投资等等。我甚至看到有人使用“人工智能寒冬”这个词。这是真正的恐惧吗？这可能是一条死胡同，还是您认为神经网络是人工智能的未来？

MARTIN FORD: I’ve seen a lot of articles in the press suggesting deep learning is being overhyped, and this hype could lead to disappointment and then less investment, and so forth. I’ve even seen the phrase “AI Winter” being used. Is that a real fear? Is this potentially a dead end, or do you think that neural networks are the future of AI?

杰弗里·辛顿：过去，当人工智能被过度炒作时——包括 20 世纪 80 年代的反向传播——人们期望它能做出伟大的事情，但实际上它并没有像他们希望的那样出色。如今，它已经做出了伟大的事情，所以不可能都是炒作。它是手机识别语音的方式，它是计算机识别照片中事物的方式，它是谷歌进行机器翻译的方式。炒作意味着你做出了巨大的承诺，但你不会兑现这些承诺，但如果你已经实现了这些承诺，那显然就不是炒作了。

GEOFFREY HINTON: In the past when AI has been overhyped—including backpropagation in the 1980s—people were expecting it to do great things, and it didn’t actually do things as great as they hoped. Today, it’s already done great things, so it can’t possibly all be just hype. It’s how your cell phone recognizes speech, it’s how a computer can recognize things in photos, and it’s how Google does machine translation. Hype means you’re making big promises, and you’re not going to live up to them, but if you’ve already achieved them, that’s clearly not hype.

我偶尔会在网络上看到一则广告，说这个行业将达到 19.9 万亿美元。这听起来似乎是一个相当大的数字，也可能是炒作，但这是一个价值数十亿美元的行业这一说法显然不是炒作，因为许多人已经向这个行业投入了数十亿美元，而且取得了成功。

I occasionally see an advertisement on the web that says it’s going to be a 19.9 trillion-dollar industry. That seems like rather a big number, and that might be hype, but the idea that it’s a multi-billion-dollar industry clearly isn’t hype, because multiple people have put billions of dollars into it and it’s worked for them.

马丁·福特：您认为未来的最佳策略是继续专注于神经网络吗？有些人仍然相信符号人工智能，他们认为可能需要一种结合深度学习和更传统方法的混合方法。您会接受这种做法吗？还是您认为该领域应该只关注神经网络？

MARTIN FORD: Do you believe the best strategy going forward is to continue to invest exclusively in neural networks? Some people still believe in symbolic AI, and they think there’s potentially a need for a hybrid approach that incorporates both deep learning and more traditional approaches. Would you be open to that, or do you think the field should focus only on neural networks?

杰弗里·辛顿：我认为，大量神经活动相互作用是大脑的工作方式，也是人工智能的工作方式。我们当然应该尝试弄清楚大脑如何进行推理，但我认为与其他事情相比，这将会来得相当晚。

GEOFFREY HINTON: I think big vectors of neural activities interacting with each other is how the brain works, and it’s how AI is going to work. We should definitely try and figure out how the brain does reasoning, but I think that’s going to come fairly late compared with other things.

我不认为混合动力系统是答案。让我们用汽车行业来打个比方。汽油发动机有一些优点，比如你可以在一个小油箱里携带很多能量，但汽油发动机也有一些非常糟糕的地方。还有电动机，与汽油发动机相比，电动机有很多优点。汽车行业的一些人同意电动发动机正在取得进展，然后说他们会制造一个混合动力系统，用电动机把汽油喷入发动机。这就是传统人工智能领域的人的想法。他们不得不承认深度学习正在做着了不起的事情，他们想把深度学习当作一种低级仆人，为他们提供使符号推理发挥作用所需的东西。这只是试图坚持他们已有的观点，而没有真正理解他们正在被彻底颠覆。

I don’t believe hybrid systems are the answer. Let’s use the car industry as an analogy. There are some good things about a petrol engine, like you can carry a lot of energy in a small tank, but there are also some really bad things about petrol engines. Then there are electric motors, which have a lot to be said in their favor compared with petrol engines. Some people in the car industry agreed that electrical engines were achieving progress and then said they’d make a hybrid system and use the electric motor to inject the petrol into the engine. That’s how people in conventional AI are thinking. They have to admit that deep learning is doing amazing things, and they want to use deep learning as a kind of low-level servant to provide them with what they need to make their symbolic reasoning work. It’s just an attempt to hang on to the view they already have, without really comprehending that they’re being swept away.

马丁·福特：从该领域的未来角度考虑，我知道您的最新项目是胶囊，我相信它是受到大脑柱状结构的启发。您是否认为研究大脑并从中获取信息并将这些见解融入您正在研究的神经网络中很重要？

MARTIN FORD: Thinking more in terms of the future of the field, I know your latest project is something you’re calling Capsules, which I believe is inspired by the columns in the brain. Do you feel that it’s important to study the brain and be informed by that, and to incorporate those insights into what you’re doing with neural networks?

杰弗里·辛顿：胶囊是六种不同想法的结合，它既复杂又具有推测性。到目前为止，它取得了一些小成功，但不能保证一定有效。现在详细谈论这一点可能还为时过早，但它的确受到了大脑的启发。

GEOFFREY HINTON: Capsules is a combination of half a dozen different ideas, and it’s complicated and speculative. So far, it’s had some small successes, but it’s not guaranteed to work. It’s probably too early to talk about that in detail, but yes, it is inspired by the brain.

当人们谈论在神经网络中使用神经科学时，大多数人对科学的认识都很幼稚。如果你想了解大脑，就会有一些基本原则，也会有很多细节。我们追求的是基本原则，如果我们使用不同类型的硬件，我们预计所有细节都会有很大的不同。图形处理器单元 (GPU) 中的硬件与大脑中的硬件非常不同，人们可能会预料到会有很多差异，但我们仍然可以寻找原则。原则的一个例子是，你大脑中的大部分知识都来自学习，而不是来自人们告诉你然后你将其存储为事实的事实。

When people talk about using neuroscience in neural networks, most people have a very naive idea of science. If you’re trying to understand the brain, there’s going to be some basic principles, and there’s going to be a whole lot of details. What we’re after is the basic principles, and we expect the details all to be very different if we use different kinds of hardware. The hardware we have in graphics processor units (GPUs) is very different from the hardware in the brain, and one might expect lots of differences, but we can still look for principles. An example of a principle is that most of the knowledge in your brain comes from learning, it doesn’t come from people telling you facts that you then store as facts.

对于传统的人工智能，人们认为你拥有这个庞大的事实数据库。你也有一些推理规则。如果我想给你一些知识，我所做的就是用某种语言表达这些事实之一，然后将其移植到你的头脑中，现在你就有了知识。这与神经网络中发生的事情完全不同：你的头脑中有很多参数，即神经元之间的连接权重，而我的头脑中有很多神经元之间的连接权重，你无法告诉我你的连接强度。无论如何，它们对我没有任何用处，因为我的神经网络与你的并不完全相同。你必须以某种方式传达有关你如何工作的信息，以便我可以以相同的方式工作，你通过给我输入和输出的示例来做到这一点。

With conventional AI, people thought that you have this big database of facts. You also have some rules of inference. If I want to give you some knowledge, what I do is simply express one of these facts in some language and then transplant it into your head, and now you have the knowledge. That’s completely different from what happens in neural networks: You have a whole lot of parameters in your head, that is weights of connections between neurons, and I have a whole lot of weights of connections between the neurons in my head, and there’s no way that you can give me your connection strengths. Anyway, they wouldn’t be any use to me because my neural network’s not exactly the same as yours. What you have to do is somehow convey information about how you are working so that I can work the same way, and you do that by giving me examples of inputs and outputs.

例如，如果你看到唐纳德·特朗普的一条推文，认为特朗普是在传达事实就大错特错了。他不是这么做的。特朗普是在说，在特定情况下，你可以选择这样一种回应方式。特朗普的追随者可以看到这种情况，他们可以看到特朗普认为他们应该如何回应，他们可以学习以与特朗普相同的方式回应。这并不是说特朗普向追随者传达了某种主张，而是通过例子传达了一种对事物的反应方式。这与拥有大量事实的系统非常不同，你可以将事实从一个系统复制到另一个系统。

For example, if you look at a tweet from Donald Trump, it’s a big mistake to think that what Trump is doing is conveying facts. That’s not what he’s doing. What Trump is doing is saying that given a particular situation, here’s a way you might choose to respond. A Trump follower can then see the situation, they can see how Trump thinks they ought to respond, and they can learn to respond the same way as Trump. It’s not that some proposition is being conveyed from Trump to the follower, it’s that a way of reacting to things has been conveyed by example. That’s very different from a system that has a big store of facts, and you can copy facts from one system to another.

马丁·福特：深度学习的绝大多数应用是否严重依赖标记数据或所谓的监督学习，而我们仍然需要解决无监督学习的问题？

MARTIN FORD: Is it true that the vast majority of applications of deep learning rely heavily on labeled data, or what’s called supervised learning, and that we still need to solve unsupervised learning?

GEOFFREY HINTON：这不完全正确。人们非常依赖标记数据，但在什么算作标记数据方面有一些微妙之处。例如，如果我给你一大串文本，并让你试着预测下一个单词，那么我会使用下一个单词作为标签，根据前面的单词判断正确答案是什么。从这个意义上说，它是有标签的，但我不需要在数据之外再加一个额外的标签。如果我给你一张图片，你想识别猫，那么我需要给你一个标签“猫”，而“猫”这个标签不是图片的一部分。我必须创建这些额外的标签，这是一项艰苦的工作。

GEOFFREY HINTON: That’s not entirely true. There’s a lot of reliance on labeled data, but there are some subtleties in what counts as labeled data. For example, if I give you a big string of text and I ask you to try and predict the next word, then I’m using the next word as a label of what the right answer is, given the previous words. In that sense, it’s labeled, but I didn’t need an extra label over and above the data. If I give you an image and you want to recognize cats, then I need to give you a label “cat,” and the label “cat” is not part of the image. I’m having to create these extra labels, and that’s hard work.

如果我只是想预测接下来会发生什么，这就是监督学习，因为接下来发生的事情充当标签，但我不需要添加额外的标签。未标记数据和标记数据之间存在着一种东西，即预测接下来会发生什么。

If I’m just trying to predict what happens next, that’s supervised learning because what happens next acts as the label, but I don’t need to add extra labels. There’s this thing in between unlabeled data and labeled data, which is predicting what comes next.

马丁·福特：如果你观察一下孩子的学习方式，就会发现他们大多是在环境中随意探索，以无人监督的方式进行学习。

MARTIN FORD: If you look at the way a child learns, though, it’s mostly wandering around the environment and learning in a very unsupervised way.

杰弗里·辛顿：回到我刚才说的，孩子在环境中四处游荡，试图预测接下来会发生什么。然后，当接下来发生的事情出现时，该事件被标记以告诉它是否正确。关键是，对于“监督”和“无监督”这两个术语，不清楚如何将它们应用于预测接下来会发生什么。

GEOFFREY HINTON: Going back to what I just said, the child is wandering around the environment trying to predict what happens next. Then when what happens next comes along, that event is labeled to tell it whether it got it right or not. The point is, with both those terms, “supervised” and “unsupervised,” it’s not clear how you apply them to predicting what happens next.

监督学习有一个很明显的例子，即我给你一张图片，并给你一个标签“猫”，然后你必须说这是一只猫；非监督学习也有一个很明显的例子，即我给你一堆图片，你必须构建图片中发生的事情的表示。最后，还有一种情况不能简单地归为这两种情况，即我给你一系列图片，你必须预测下一张图片。在这种情况下，你不清楚应该称之为监督学习还是非监督学习，这会引起很多困惑。

There’s a nice clear case of supervised learning, which is that I give you an image and I give you the label “cat,” then you have to say it’s a cat, then there’s a nice clear case of unsupervised learning, which is if I give you a bunch of images, and you have to build representations of what’s going on in the images. Finally, there’s something that doesn’t fall simply into either camp, which is if I give you a sequence of images and you have to predict the next image. It’s not clear in that case whether you should call that supervised learning or unsupervised learning, and that causes a lot of confusion.

马丁·福特：您是否认为解决无监督学习的一般形式是需要克服的主要障碍之一？

MARTIN FORD: Would you view solving a general form of unsupervised learning as being one of the primary obstacles that needs to be overcome?

杰弗里·辛顿：是的。但从这个意义上讲，无监督学习的一种形式是预测接下来会发生什么，我的观点是，你可以应用监督学习算法来做到这一点。

GEOFFREY HINTON: Yes. But in that sense, one form of unsupervised learning is predicting what happens next, and my point is that you can apply supervised learning algorithms to do that.

马丁·福特：您如何看待 AGI？您如何定义它？我认为它是指人类级别的人工智能，即能够像人类一样进行一般推理的人工智能。这是您的定义吗？还是您认为它是其他定义？

MARTIN FORD: What do you think about AGI, and how would you define that? I would take it to mean human-level artificial intelligence, namely an AI that can reason in a general way, like a human. Is that your definition, or would you say it’s something else?

GEOFFREY HINTON：我对这个定义很满意，但我认为人们对未来会是什么样子有各种各样的假设。人们认为我们将会得到越来越聪明的人工智能，但我认为这种说法有两点错误。其一，深度学习或神经网络在某些方面会比我们做得好得多，但在其他方面它们仍然比我们差很多。它们并不是在所有方面都会变得一致好。例如，它们在解读医学图像方面会好得多，但在推理方面仍然差很多。从这个意义上说，人工智能的发展不会是一致的。

GEOFFREY HINTON: I’m happy with that definition, but I think people have various assumptions of what the future’s going to look like. People think that we’re going to get individual AIs that get smarter and smarter, but I think there are two things wrong with that picture. One is that deep learning, or neural networks are going to get much better than us at some things, while they’re still quite a lot worse than us at other things. It’s not like they’re going to get uniformly better at everything. They’re going to be much better, for example, at interpreting medical images, while they’re still a whole lot worse at reasoning about them. In that sense, it’s not going to be uniform.

第二个错误是，人们总是将其视为单个人工智能，而忽略了其社会性。仅从计算角度来看，制造非常先进的智能将涉及创建智能系统社区，因为社区比单个系统可以看到更多的数据。如果这完全是为了查看大量数据，那么我们将不得不将这些数据分布在许多不同的智能系统中，并让它们相互通信，这样它们之间，作为一个社区，就可以从大量数据中学习，这意味着在未来，它的社区性将至关重要。

The second thing that’s wrong is that people always think about it as individual AIs, and they ignore the social aspect of it. Just for pure computational reasons, making very advanced intelligence is going to involve making communities of intelligent systems because a community can see much more data than an individual system. If it’s all a question of seeing a lot of data, then we’re going to have to distribute that data across lots of different intelligent systems and have them communicate with one another so that between them, as a community, they can learn from a huge amount of data meaning that in the future, the community aspect of it is going to be essential.

马丁·福特：您是否认为它是互联网上互联智能的一个新兴特性？

MARTIN FORD: Do you envision it as being an emergent property of connected intelligences on the internet?

杰弗里·辛顿：不，人也是一样。你之所以知道大部分知识，并不是因为你自己从数据中提取了这些信息，而是因为其他人多年来已经从数据中提取了信息。然后他们给你提供训练经验，让你无需从数据中进行原始提取就能获得相同的理解。我认为人工智能也会如此。

GEOFFREY HINTON: No, it’s the same with people. The reason that you know most of what you know is not because you yourself extracted that information from data, it’s because other people, over many years, have extracted information from data. They then gave you training experiences that allowed you to get to the same understanding without having to do the raw extraction from data. I think it’ll be like that with artificial intelligence too.

马丁·福特：您认为 AGI（无论是单个系统还是一组相互作用的系统）是否可行？

MARTIN FORD: Do you think AGI, whether it’s an individual system or a group of systems that interact, is feasible?

杰弗里·辛顿：哦，是的。我的意思是 OpenAI 已经开发出了一种可以作为一个团队玩相当复杂的计算机游戏的东西。

GEOFFREY HINTON: Oh, yes. I mean OpenAI already has something that plays quite sophisticated computer games as a team.

马丁·福特：您认为什么时候人工智能或者一组人工智能能够拥有与人类相同的推理、智能和能力？

MARTIN FORD: When do you think it might be feasible for an artificial intelligence, or a group of AIs that come together, to have the same reasoning, intelligence, and capability as a human being?

杰弗里·辛顿：如果你去推理，我认为这将是我们以后真正擅长的事情之一，但大型神经网络在推理方面真正达到与人类一样优秀还需要相当长的时间。话虽如此，在我们达到这一点之前，它们在各种其他事情上都会表现得更好。

GEOFFREY HINTON: If you go for reasoning, I think that’s going to be one of the things we get really good at later on, but it’s going to be quite a long time before big neural networks are really as good as people at reasoning. That being said, they’ll be better at all sorts of other things before we get to that point.

马丁·福特：那么，对于整体通用人工智能来说，计算机系统的智能是否能与人类一样好呢？

MARTIN FORD: What about for a holistic AGI, though, where a computer system’s intelligence is as good as a person?

杰弗里·辛顿：我认为，有一种预设认为，人工智能的发展方式是制造出像《星际迷航》中那样的通用机器人。如果你的问题是“我们什么时候才能拥有指挥官 Data？”，那么我认为事情不会这样发展。我认为我们不会得到那种单一的通用的东西。我还认为，就通用推理能力而言，这在相当长的一段时间内都不会发生。

GEOFFREY HINTON: I think there’s a presupposition that the way AIs can develop is by making individuals that are general-purpose robots like you see on Star Trek. If your question is, “When are we going to get a Commander Data?”, then I don’t think that’s how things are going to develop. I don’t think we’re going to get single, general-purpose things like that. I also think, in terms of general reasoning capacity, it’s not going to happen for quite a long time.

马丁·福特：想象一下通过图灵测试，不是五分钟而是两个小时，这样你就可以像人类一样进行广泛的对话。这可行吗，无论是一个系统还是某个系统社区？

MARTIN FORD: Think of it in terms of passing the Turing test, and not for five minutes but for two hours, so that you can have a wide-ranging conversation that’s as good as a human being. Is that feasible, whether it’s one system or some community of systems?

杰弗里·辛顿：我认为，在未来 10 到 100 年内，这种事情发生的可能性很大。我认为可能性很小，它会在下一个十年结束前发生，而且我认为，在下一个 100 年内，人类被其他事物消灭的可能性也很大。

GEOFFREY HINTON: I think there’s a reasonable amount of probability that it will happen in somewhere between 10 and 100 years. I think there’s a very small probability, it’ll happen before the end of the next decade, and I think there’s also a big probability that humanity gets wiped out by other things before the next 100 years occurs.

马丁·福特：你的意思是通过其他生存威胁，比如核战争或瘟疫吗？

MARTIN FORD: Do you mean through other existential threats like a nuclear war or a plague?

杰弗里·辛顿：是的，我认为是这样。换句话说，我认为有两个生存威胁比人工智能大得多。一个是全球核战争，另一个是分子生物学实验室里心怀不满的研究生制造出一种极具传染性、致命性且潜伏期很长的病毒。我认为这才是人们应该担心的，而不是超智能系统。

GEOFFREY HINTON: Yes, I think so. In other words, I think there are two existential threats that are much bigger than AI. One is global nuclear war, and the other is a disgruntled graduate student in a molecular biology lab making a virus that’s extremely contagious, extremely lethal, and has a very long incubation time. I think that’s what people should be worried about, not ultra-intelligent systems.

马丁·福特：有些人，比如 DeepMind 的 Demis Hassabis，确实相信他们可以建立那种你认为不会实现的系统。你怎么看？你认为这是一项徒劳的任务吗？

MARTIN FORD: Some people, such as Demis Hassabis at DeepMind, do believe that they can build the kind of system that you’re saying you don’t think is going to come into existence. How do you view that? Do you think that it is a futile task?

杰弗里·辛顿：不，我认为这是因为我和戴米斯对未来的预测不同。

GEOFFREY HINTON: No, I view that as Demis and me having different predictions about the future.

马丁·福特：我们来谈谈人工智能的潜在风险。我写过的一个特别的挑战是它对就业市场和经济的潜在影响。您是否认为所有这些都可能引发一场新的工业革命并彻底改变就业市场？如果是这样，这是我们需要担心的事情吗？还是说这又是一个被夸大了的事情？

MARTIN FORD: Let’s talk about the potential risks of AI. One particular challenge that I’ve written about is the potential impact on the job market and the economy. Do you think that all of this could cause a new Industrial Revolution and completely transform the job market? If so, is that something we need to worry about, or is that another thing that’s perhaps overhyped?

杰弗里·辛顿：如果你能大幅提高生产力，生产出更多好东西，那应该是一件好事。但这是否是一件好事完全取决于社会制度，而根本不取决于技术。人们看待技术的方式，好像技术进步是一个问题。问题在于社会制度，以及我们是否会有一个公平分享的社会制度，还是一个将所有改进都集中在 1% 的人身上，而将其他人视为垃圾的社会制度。这与技术无关。

GEOFFREY HINTON: If you can dramatically increase productivity and make more goodies to go around, that should be a good thing. Whether or not it turns out to be a good thing depends entirely on the social system, and doesn’t depend at all on the technology. People are looking at the technology as if the technological advances are a problem. The problem is in the social systems, and whether we’re going to have a social system that shares fairly, or one that focuses all the improvement on the 1% and treats the rest of the people like dirt. That’s nothing to do with technology.

马丁·福特：然而，这个问题的出现是因为许多工作可能会被淘汰——尤其是那些可预测且容易被自动化的工作。对此，社会的一种反应是基本收入，你同意吗？

MARTIN FORD: That problem comes about, though, because a lot of jobs could be eliminated—in particular, jobs that are predictable and easily automated. One social response to that is a basic income, is that something that you agree with?

杰弗里·辛顿：是的，我认为基本收入是一个非常明智的想法。

GEOFFREY HINTON: Yes, I think a basic income is a very sensible idea.

马丁·福特：那么，您认为需要采取政策应对措施来解决这个问题吗？有些人认为我们应该顺其自然，但这也许是不负责任的。

MARTIN FORD: Do you think, then, that policy responses are required to address this? Some people take a view that we should just let it play out, but that’s perhaps irresponsible.

杰弗里·辛顿：我之所以搬到加拿大，是因为这里的税率较高，而且我认为合理的税收是好事。政府应该建立机制，这样当人们为自己的利益行事时，每个人都会从中受益。高税收就是这样一种机制：当人们变得富有时，其他人都会得到税收的帮助。我当然同意，要确保人工智能惠及每个人，还有很多工作要做。

GEOFFREY HINTON: I moved to Canada because it has a higher taxation rate and because I think taxes done right are good things. What governments ought to be is mechanisms put in place so that when people act in their own self-interest, it helps everybody. High taxation is one such mechanism: when people get rich, everybody else gets helped by the taxes. I certainly agree that there’s a lot of work to be done in making sure that AI benefits everybody.

马丁·福特：您认为人工智能还存在哪些其他风险，比如武器化？

MARTIN FORD: What about some of the other risks that you would associate with AI, such as weaponization?

杰弗里·辛顿：是的，我对普京总统最近说的一些话感到担忧。我认为人们现在应该非常积极地努力让国际社会以对待化学武器和大规模杀伤性武器同样的方式对待那些无需人员介入就能杀人的武器。

GEOFFREY HINTON: Yes, I am concerned by some of the things that President Putin has said recently. I think people should be very active now in trying to get the international community to treat weapons that can kill people without a person in the loop the same way as they treat chemical warfare and weapons of mass destruction.

马丁·福特：您是否赞成对此类研究和开发实行某种形式的暂停？

MARTIN FORD: Would you favor some kind of a moratorium on that type of research and development?

杰弗里·辛顿：你们不会暂停这类研究，就像你们没有暂停神经毒剂的研发一样，但你们确实有国际机制阻止它们被广泛使用。

GEOFFREY HINTON: You’re not going to get a moratorium on that type of research, just as you haven’t had a moratorium on the development of nerve agents, but you do have international mechanisms in place that have stopped them being widely used.

马丁·福特：除了使用军事武器之外，还有其他风险吗？还有其他问题吗，比如隐私和透明度？

MARTIN FORD: What about other risks, beyond the military weapon use? Are there other issues, like privacy and transparency?

杰弗里·辛顿：我认为利用这些数据操纵选举和选民的行为令人担忧。剑桥分析公司是由机器学习专家鲍勃·默瑟创立的，你已经看到剑桥分析公司造成了巨大的损失。我们必须认真对待此事。

GEOFFREY HINTON: I think using it to manipulate elections and to manipulate voters is worrying. Cambridge Analytica was set up by Bob Mercer who was a machine learning person, and you’ve seen that Cambridge Analytica did a lot of damage. We have to take that seriously.

马丁·福特：您认为还有监管的空间吗？

MARTIN FORD: Do you think that there’s a place for regulation?

杰弗里·辛顿：是的，有很多监管。这是一个非常有趣的问题，但我不是这方面的专家，所以没有什么可以提供的。

GEOFFREY HINTON: Yes, lots of regulation. It’s a very interesting issue, but I’m not an expert on it, so don’t have much to offer.

马丁·福特：那么，全球人工智能军备竞赛如何呢？您认为一个国家不要领先其他国家太多，这重要吗？

MARTIN FORD: What about the global arms race in general AI, do you think it’s important that one country doesn’t get too far ahead of the others?

杰弗里·辛顿：你谈论的是全球政治。长期以来，英国是一个占主导地位的国家，但他们表现不佳，后来是美国，他们表现也不佳，如果轮到中国，我预计他们也不会表现得很好。

GEOFFREY HINTON: What you’re talking about is global politics. For a long time, Britain was a dominant nation, and they didn’t behave very well, and then it was America, and they didn’t behave very well, and if it becomes the Chinese, I don’t expect them to behave very well.

马丁·福特：我们应该制定某种形式的产业政策吗？美国和其他西方政府是否应该将人工智能作为国家优先事项？

MARTIN FORD: Should we have some form of industrial policy? Should the United States and other Western governments focus on AI and make it a national priority?

杰弗里·辛顿：技术将会发生巨大的进步，如果一个国家不努力跟上这种进步，那他就太疯狂了，所以我认为，显然应该在这方面投入大量资金。这对我来说似乎是常识。

GEOFFREY HINTON: There are going to be huge technological developments, and countries would be crazy not to try and keep up with that, so obviously, I think there should be a lot of investment in it. That seems common sense to me.

马丁·福特：总体而言，您对这一切持乐观态度吗？您认为人工智能带来的好处会超过坏处吗？

MARTIN FORD: Overall, are you optimistic about all of this? Do you think that the rewards from AI are going to outweigh the downsides?

杰弗里·辛顿：我希望其好处能够大于坏处，但我不知道是否如此，这是社会制度的问题，而不是技术的问题。

GEOFFREY HINTON: I hope the rewards will outweigh the downsides, but I don’t know whether they will, and that’s an issue of social systems, not with the technology.

马丁·福特：人工智能领域人才严重短缺，每个人都在招聘。对于想要进入这个领域的年轻人，您有什么建议吗？有什么建议可以帮助吸引更多人，让他们成为人工智能和深度学习方面的专家？

MARTIN FORD: There’s an enormous talent shortage in AI and everyone’s hiring. Is there any advice you would give to a young person that wants to get into this field, anything that might help attract more people and enable them to become expert in AI and in deep learning, that you can offer?

杰弗里·辛顿：我担心可能没有足够多的人对基本原则持批评态度。Capsules 的理念是，也许我们做事的一些基本方式并不是最好的方式，我们应该撒下更大的网。我们应该考虑一些替代我们所做的一些最基本的假设的方法。我给人们的一个建议是，如果你直觉地认为人们所做的事情是错误的，并且可能有更好的事情，那么你应该遵循你的直觉。

GEOFFREY HINTON: I’m worried that there may not be enough people who are critical of the basics. The idea of Capsules is to say, maybe some of the basic ways we’re doing things aren’t the best way of doing things, and we should throw a wider net. We should think about alternatives to some of the very basic assumptions we’re making. The one piece of advice I give people is that if you have intuitions that what people are doing is wrong and that there could be something better, you should follow your intuitions.

你很可能是错的，但除非人们遵循直觉，彻底改变现状，否则我们就会陷入困境。一个令人担心的是，我认为真正新想法的最丰富来源是大学里得到良好建议的研究生。他们有自由提出真正新的想法，他们学到的东西足够多，所以他们不会只是重复历史，我们需要保持这一点。攻读硕士学位然后直接进入行业的人不会提出全新的想法。我认为你需要坐下来思考几年。

You’re quite likely to be wrong, but unless people follow the intuitions when they have them about how to change things radically, we’re going to get stuck. One worry is that I think the most fertile source of genuinely new ideas is graduate students being well advised in a university. They have the freedom to come up with genuinely new ideas, and they learn enough so that they’re not just repeating history, and we need to preserve that. People doing a master’s degree and then going straight into the industry aren’t going to come up with radically new ideas. I think you need to sit and think for a few years.

马丁·福特：加拿大似乎是一个深度学习融合中心。这是偶然的，还是加拿大有什么特殊之处促成了这一发展？

MARTIN FORD: There seems to be a hub of deep learning coalescing in Canada. Is that just random, or is there something special about Canada that helped with that?

GEOFFREY HINTON：加拿大高等研究院 (CIFAR) 为高风险领域的基础研究提供资金，这一点非常重要。另外，幸运的是，曾担任我博士后研究的 Yann LeCun 和 Yoshua Bengio 都在加拿大。我们三人能够形成富有成效的合作关系，加拿大高等研究院也为这次合作提供了资金。当时，我们所有人都在一个相当恶劣的环境中有点孤立 — — 直到最近，深度学习的环境才变得相当恶劣 — — 获得这笔资金非常有帮助，它让我们有相当多的时间在小型会议上相处，我们可以真正分享未发表的想法。

GEOFFREY HINTON: The Canadian Institute for Advanced Research (CIFAR) provided funding for basic research in high-risk areas, and that was very important. There’s also a lot of good luck in that both Yann LeCun, who was briefly my postdoc, and Yoshua Bengio were also in Canada. The three of us could form a collaboration that was very fruitful, and the Canadian Institute for Advanced Research funded that collaboration. This was at a time when all of us would have been a bit isolated in a fairly hostile environment—the environment for deep learning was fairly hostile until quite recently—it was very helpful to have this funding that allowed us to spend quite a lot of time with each other in small meetings, where we could really share unpublished ideas.

马丁·福特：那么，这是加拿大政府为保持深度学习活力而采取的一项战略投资？

MARTIN FORD: So, it was a strategic investment on the part of the Canadian government to keep deep learning alive?

杰弗里·辛顿：是的。基本上，加拿大政府每年投入 50 万美元对先进的深度学习进行大量投资，对于即将发展成为价值数十亿美元的产业来说，这种投入相当高效。

GEOFFREY HINTON: Yes. Basically, the Canadian government is significantly investing in advanced deep learning by spending half a million dollars a year, which is pretty efficient for something that’s going to turn into a multi-billion-dollar industry.

马丁·福特：说到加拿大人，你和同事乔丹·彼得森有过交流吗？多伦多大学似乎出现了各种各样的混乱……

MARTIN FORD: Speaking of Canadians, do you have any interaction with your fellow faculty member, Jordan Peterson? It seems like there’s all kinds of disruption coming out of the University of Toronto...

杰弗里·辛顿：哈哈！好吧，我只能说，他是一个不知道何时该闭嘴的人。

GEOFFREY HINTON: Ha! Well, all I’ll say about that is that he’s someone who doesn’t know when to keep his mouth shut.

杰弗里·辛顿 于 1978 年获得剑桥大学国王学院的学士学位，并于爱丁堡大学获得人工智能博士学位。在卡内基梅隆大学担任教职五年后，他成为加拿大高等研究院的研究员，并转至多伦多大学计算机科学系，目前是该系的名誉杰出教授。他还是谷歌的副总裁兼工程研究员，以及人工智能向量研究所的首席科学顾问。

GEOFFREY HINTON received his undergraduate degree from Kings College, Cambridge and his PhD in Artificial Intelligence from the University of Edinburgh in 1978. After five years as a faculty member at Carnegie-Mellon University, he became a fellow of the Canadian Institute for Advanced Research and moved to the Department of Computer Science at the University of Toronto where he is now an Emeritus Distinguished Professor. He is also a Vice President & Engineering Fellow at Google and Chief Scientific Adviser of the Vector Institute for Artificial Intelligence.

Geoff 是引入反向传播算法的研究人员之一，也是第一个使用反向传播来学习词向量的研究人员。他对神经网络研究的其他贡献包括玻尔兹曼机、分布式表示、时间延迟神经网络、专家混合、变分学习和深度学习。他在多伦多的研究小组在深度学习方面取得了开创性的突破，彻底改变了语音识别和对象分类。

Geoff was one of the researchers who introduced the backpropagation algorithm and the first to use backpropagation for learning word embeddings. His other contributions to neural network research include Boltzmann machines, distributed representations, time-delay neural nets, mixtures of experts, variational learning and deep learning. His research group in Toronto made seminal breakthroughs in deep learning that revolutionized speech recognition and object classification.

Geoff 是英国皇家学会院士、美国国家工程院外籍院士和美国艺术与科学学院外籍院士，获奖包括 David E. Rumelhart 奖、IJCAI 研究卓越奖、Killam 工程奖、IEEE Frank Rosenblatt 奖章、IEEE James Clerk Maxwell 金奖、NEC C&C 奖、BBVA 奖，以及加拿大科学与工程领域最高奖项 NSERC Herzberg 金奖。 ”

Geoff is a fellow of the UK Royal Society, a foreign member of the US National Academy of Engineering and a foreign member of the American Academy of Arts and Sciences. His awards include the David E. Rumelhart prize, the IJCAI award for research excellence, the Killam prize for Engineering, the IEEE Frank Rosenblatt medal, the IEEE James Clerk Maxwell Gold medal, the NEC C&C award, the BBVA award, and the NSERC Herzberg Gold Medal, which is Canada’s top award in science and engineering.”

牛津大学教授、人类未来研究所所长

PROFESSOR, UNIVERSITY OF OXFORD AND DIRECTOR OF THE FUTURE OF HUMANITY INSTITUTE

尼克·博斯特罗姆被公认为超级智能领域的世界顶级专家之一，他研究人工智能和机器学习可能给人类带来的生存风险。他是牛津大学人类未来研究所的创始主任，该研究所是一个多学科研究机构，研究人类及其前景的宏观问题。他著述颇丰，出版了 200 多部作品，包括 2014 年《纽约时报》畅销书《超级智能：路径、危险、策略》。

Nick Bostrom is widely recognized as one of the world’s top experts on superintelligence and the existential risks that AI and machine learning could potentially pose for humanity. He is the Founding Director of the Future of Humanity Institute at the University of Oxford, a multidisciplinary research institute studying big-picture questions about humanity and its prospects. He is a prolific author of over 200 publications, including the 2014 New York Times bestseller Superintelligence: Paths, Dangers, Strategies.

马丁·福特：您曾经写过关于创建超级智能的风险的文章——当 AGI 系统将其精力转向自我改进时，可能会出现超级智能，从而形成一个递归改进循环，最终产生远远优于人类的智能。

MARTIN FORD: You’ve written about the risks of creating a superintelligence—an entity that could emerge when an AGI system turns its energies toward improving itself, creating a recursive improvement loop that results in an intelligence that is vastly superior to humans.

尼克·博斯特罗姆：是的，这是一个场景和一个问题，但是向机器智能时代的转变还可能出现其他场景和其他方式，而且肯定还存在其他需要人们担心的问题。

NICK BOSTROM: Yes, that’s one scenario and one problem, but there are other scenarios and other ways this transition to a machine intelligence era could unfold, and there are certainly, other problems one could be worried about.

马丁·福特：您特别关注的一个想法是控制或协调问题，即机器智能的目标或价值观可能会导致对人类有害的结果。您能用通俗易懂的语言更详细地解释一下协调问题或控制问题是什么吗？

MARTIN FORD: One idea you’ve focused on especially is the control or alignment problem where a machine intelligence’s goals or values might result in outcomes that are harmful to humanity. Can you go into more detail on what that alignment problem, or control problem, is in layman’s terms?

尼克·博斯特罗姆：与其他技术不同的是，非常先进的人工智能系统有一个显著的问题，那就是它不仅存在人类滥用技术的可能性（当然，这也是我们在其他技术中看到的），而且还有技术滥用自身的可能性。换句话说，你创造了一个人工智能代理或一个过程，它有自己的目标和目的，而且它非常有能力实现这些目标，因为在这种情况下，它是超级智能。令人担忧的是，这个强大的系统试图优化的目标与我们的人类价值观不同，甚至可能与我们想要在这个世界上实现的目标背道而驰。那么，如果人类试图实现一件事，而超级智能系统试图实现另一件事，那么超级智能很可能会获胜并得偿所愿。

NICK BOSTROM: Well, one distinctive problem with very advanced AI systems that’s different from other technologies is that it presents not only the possibility of humans misusing the technology—that’s something we see with other technologies, of course—but also the possibility that the technology could misuse itself, as it were. In other words, you create an artificial agent or a process that has its own goals and objectives, and it is very capable of achieving those objectives because, in this scenario, it is superintelligent. The concern is that the objectives that this powerful system is trying to optimize for are different from our human values, and maybe even at cross-purposes with what we want to achieve in the world. Then if you have humans trying to achieve one thing and a superintelligent system trying to achieve something different, it might well be that the superintelligence wins and gets its way.

我们担心的不是它会因为奴役它而憎恨或怨恨我们，或者它会突然产生意识并反抗我们，而是它会非常有效地追求与我们真正想要的目标不同的目标。然后你会看到一个按照外星人标准塑造的未来。那么控制问题或协调问题就是如何设计人工智能系统，使它们成为人类意志的延伸？从某种意义上说，我们用我们的意图塑造它们的行为，而不是随机、不可预见和不想要的目标突然出现？

The concern is not that it would hate or resent us for enslaving it, or that suddenly a spark of consciousness would arise and it would rebel, but rather that it would be very competently pursuing an objective that differs from what we really want. Then you get a future shaped in accordance with alien criteria. The control problem, or the alignment problem, then is how do you engineer AI systems so that they are an extension of human will? In the sense that we have our intentions shape their behavior as opposed to a random, unforeseen and unwanted objective cropping up there?

马丁·福特：你举了一个著名的例子，一个生产回形针的系统。这个想法是，当一个系统被构思出来并被赋予一个目标时，它会以超级智能的能力去追求这个目标，但它做这件事的方式并没有考虑到常识，所以它最终会伤害到我们。你举的例子是一个系统把整个宇宙变成回形针，因为它是一个回形针优化器。这是对对齐问题的一个很好的阐述吗？

MARTIN FORD: You have a famous example of a system that manufactures paperclips. The idea is that when a system is conceived and given an objective, it pursues that goal with a superintelligent competence, but it does it in a way that doesn’t consider common sense, so it ends up harming us. The example you give is a system that turns the whole universe into paperclips because it’s a paperclip optimizer. Is that a good articulation of the alignment problem?

尼克·博斯特罗姆：回形针的例子代表了一类更广泛的可能失败情况，即你要求一个系统做一件事，也许一开始结果还不错，但后来它突然得出了一个我们无法控制的结论。这是一个卡通例子，你设计一个人工智能来运营一个回形针工厂。它一开始很笨，但它越聪明，就能把回形针工厂运营得越好，这家工厂的老板非常高兴，想要取得更多进展。然而，当人工智能变得足够聪明时，它会意识到还有其他方法可以让世界上的回形针数量更多，这可能涉及夺走人类的控制权，将整个星球变成回形针或太空探测器，可以出去把宇宙变成更多的回形针。

NICK BOSTROM: The paperclip example is a stand-in for a wider category of possible failures where you ask a system to do one thing and, perhaps, initially things turn out pretty well but then it races to a conclusion that is beyond our control. It’s a cartoon example, where you design an AI to operate a paperclip factory. It’s dumb initially, but the smarter it gets, the better it operates the paperclip factory, and the owner of this factory is very pleased and wants to make more progress. However, when the AI becomes sufficiently smart, it realizes that there are other ways of achieving an even greater number of paperclips in the world, which might then involve taking control away from humans and indeed turning the whole planet into paperclips or into space probes that can go out and transform the universe into more paperclips.

这里的要点是，你可以用回形针代替几乎任何其他目标，如果你仔细想想在这个世界上真正最大化这个目标意味着什么，你就会发现，除非你真的非常小心地指定你的目标，否则，最大化实现该目标的副作用是，人类和我们关心的事物将被消灭。

The point here is that you could substitute almost any other goal you want for paperclips and if you think through what it would mean for that goal to be truly maximized in this world, that unless you’re really, really careful about how you specify your goals, you will find that as a side effect of maximizing for that goal human beings and the things we care about would be stamped out.

马丁·福特：当我听到有人描述这个问题时，总是说我们给系统设定了一个目标，但它却以一种我们并不满意的方式去实现这个目标。然而，我从未听说过有哪个系统会简单地改变它的目标，我也不太明白为什么这不是一个值得担心的问题。为什么超级智能系统不能在某个时候决定拥有不同的目标或目的呢？人类一直都是这样做的！

MARTIN FORD: When I hear this problem described, it’s always given as a situation where we give the system a goal, and then it pursues that goal in a way that we’re not happy with. However, I never hear of a system that simply changes its goal, and I don’t quite understand why that is not a concern. Why couldn’t a superintelligent system at some point just decide to have different goals or objectives? Humans do it all of the time!

尼克·博斯特罗姆：这个问题之所以不那么令人担忧，是因为尽管超级智能有能力改变其目标，但你必须考虑它选择目标的标准。它会根据当时的目标做出选择。在大多数情况下，智能体改变目标是一个非常糟糕的战略举措，因为它可以预测，未来将不会有智能体追求其当前目标，而是会有一个智能体追求其他目标。这往往会产生比其当前目标排名更低的结果，而根据定义，当前目标正是它选择行动的标准。因此，一旦你拥有一个足够复杂的推理系统，你就会期望它能够解决这个问题，从而能够实现内部目标稳定性。

NICK BOSTROM: The reason why this seems less of a concern is that although a superintelligence would have the ability to change its goals, you have to consider the criteria it uses to choose its goals. It would make that choice based on the goals it has at that moment. In most situations, it would be a very poor strategic move for an agent to change its goals because it can predict that in the future, there will then not be an agent pursuing its current goal but instead an agent pursuing some different goal. This would tend to produce outcomes that would rank lower by its current goals, which by definition here are what it is using as the criteria by which to select actions. So, once you have a sufficiently sophisticated reasoning system, you expect it to figure this out and therefore be able to achieve internal goal stability.

人类是一团乱麻。我们没有一个特定的目标，而我们所追求的所有其他目标都是子目标。我们的大脑中有不同的部分在朝着不同的方向发展，如果你增加我们的激素水平，我们会突然改变这些价值观。人类不像机器那样稳定，也许没有一个非常清晰、紧凑的目标最大化代理描述。这就是为什么我们人类有时会决定改变我们的目标。这并不是我们决定改变我们的目标；而是我们的目标在变化。或者，我们所说的“目标”并不是指我们判断事物的基本标准，而是指某个特定的目标，当然，它会随着情况的变化或我们发现新的计划而改变。

Humans are a mess. We don’t have a particular goal from which all the other objectives we pursue are sub-goals. We have different parts of our minds that are pulling in different directions, and if you increase our hormone levels, we suddenly change those values. Humans are not stable in the same way as machines, and maybe don’t have a very clean, compact description as goal-maximizing agents. That’s why it can seem that we humans sometimes decide to change our goals. It’s not so much us deciding to change our goals; it’s our goals just changing. Alternatively, by “goals,” we don’t mean our fundamental criteria for judging things, but just some particular objective, which of course can change as circumstances change or we discover new plans.

马丁·福特：许多这方面的研究都是基于神经科学的，因此，将人类大脑的想法注入机器智能中是可行的。想象一下，一个超级智能可以掌握人类的所有知识。它将能够阅读整个人类历史。它将阅读有关有权势的人以及他们有不同的目的和目标的资料。机器也可能受到病理的影响。人类大脑有各种各样的问题，有些药物可以改变大脑的工作方式。我们怎么知道机器空间中没有类似的东西呢？

MARTIN FORD: A lot of the research going into this is informed by neural science, though, so there are ideas coming from the human brain being injected into machine intelligence. Imagine a superintelligence that has at its disposal all of human knowledge. It would be able to read all of human history. It would read about powerful individuals, and how they had different objectives and goals. The machine could also conceivably be subject to pathologies. The human brain has all kinds of problems, and there are drugs that can change the way the brain works. How do we know there’s not something comparable in the machine space?

尼克·博斯特罗姆：我认为，特别是在开发早期阶段，在机器充分理解人工智能的工作原理之前，在机器能够自我调整而不弄乱自己之前，这种情况是有可能发生的。归根结底，开发技术是为了防止目标被破坏，这背后有着共同的工具性原因。我希望一个足够强大的系统能够开发这些技术来实现目标稳定性，事实上，它可能会优先开发这些技术。然而，如果它很着急，或者它的能力还不够强——如果它的能力大致与人类相当——那么事情就有可能变得混乱。人们可能会实施一项变革，希望它能成为一个更有效的思考者，但结果却在改变其目标函数方面产生了一些副作用。

NICK BOSTROM: I think there well could be, particularly in the earlier stages of development, before the machine achieves sufficient understanding of how AI works to be able to modify itself without messing itself up. Ultimately, there are convergent instrumental reasons for developing technology to prevent your goals from being corrupted. I would expect a sufficiently capable system to develop those technologies for goal stability, and indeed it might place some priority on developing them. However, if it’s in a rush or if it’s not yet very capable—if it’s roughly at the human level—the possibility certainly exists that things could get scrambled. A change might be implemented with the hope that it would maybe make it a more effective thinker, but it turns out to have some side effect in changing its objective function.

马丁·福特：我担心的另一件事是，我们总是担心机器不会按照我们的意愿行事，这里的“我们”指的是人类集体，好像人类有某种普遍的欲望或价值观。然而，如果你看看今天的世界，事实并非如此。世界有不同的文化，有不同的价值观。在我看来，第一个机器智能在哪里开发可能很重要。把机器和全人类视为一个整体是否幼稚？在我看来，事情似乎比这要复杂得多

MARTIN FORD: The other thing that I worry about is that it’s always a concern about how the machine is not going to do what we want, where “we” applies to collective humanity as though there’s some sort of universal set of human desires or values. Yet, if you look at the world today, that’s really not the case. The world has different cultures with different value sets. It seems to me that it might matter quite a lot where the first machine intelligence is developed. Is it naive to talk about the machine and all of humanity as being one entity? To me, it just seems like things are a lot messier than that

尼克·博斯特罗姆：你试图将大问题分解成小问题，然后逐步解决。你试图分解出整体挑战的一个组成部分，在这种情况下，这就是技术问题，即如何实现人工智能与人类价值观的一致，让机器按照开发人员的意愿行事。除非你找到解决方案，否则你甚至没有权利尝试解决更广泛的政治问题，即确保我们人类将把这项强大的技术用于某些有益的目的。

NICK BOSTROM: You try to break up the big problem into smaller problems in order then to make progress on them. You try to break out one component of the overall challenge, in this case that is the technical problem of how to achieve AI alignment with any human values to get the machine to do what its developers want it to do. Unless you have a solution to that, you don’t have the privilege even to try for a solution to the wider, political problems of ensuring that we humans will then use this powerful technology for some beneficial purpose.

你需要解决技术问题，才能有机会争论谁的价值观，或者不同价值观在多大程度上应该指导这项技术的使用。当然，即使你找到了技术控制问题的解决方案，你也只是解决了整个挑战的一部分。你还需要想出一种方法，让我们能够和平地、造福全人类地使用它。

You need to solve the technical problem to get the opportunity to squabble over whose values, or in what degrees different values should guide the use of this technology. It is true, of course, that even if you have a solution to the technical control problem, you’ve really only solved part of the overall challenge. You also then need to figure out a way that we can use this peacefully and in a way that benefits all of humanity.

马丁·福特：解决这个技术控制问题，即如何制造一台与目标保持一致的机器，是你在人类未来研究所正在研究的内容吗？其他智库如 OpenAI 和机器智能研究所也在关注什么？

MARTIN FORD: Is solving that technical control problem, in terms of how to build a machine that remains aligned with the objective, what you’re working on at the Future of Humanity Institute, and what other think tanks like OpenAI and the Machine Intelligence Research Institute are focusing on?

尼克·博斯特罗姆：是的，没错。我们确实有一个小组在研究这个问题，但我们也在研究其他事情。我们还有一个人工智能治理小组，专注于与机器智能进步相关的治理问题。

NICK BOSTROM: Yes, that’s right. We do have a group working on that, but we’re also working on other things. We also have a governance of AI group, that is focused on the governance problems related to advances in machine intelligence.

马丁·福特：您认为像您这样的智库对于人工智能治理而言是适当的资源配置吗？或者您认为政府应该更大规模地参与其中？

MARTIN FORD: Do you think that think tanks like yours are an appropriate level of resource allocation for AI governance, or do you think that governments should jump into this at a larger scale?

尼克·博斯特罗姆：我认为人工智能安全方面可以投入更多资源。实际上，这不仅仅是我们：DeepMind 也有一个与我们合作的人工智能安全小组，但我确实认为更多的资源是有益的。现在的人才和资金已经比四年前多得多了。从百分比来看，这是一个快速增长的轨迹，尽管从绝对值来看，这仍然是一个非常小的领域。

NICK BOSTROM: I think there could be more resources on AI safety. It’s not actually just us: DeepMind also has an AI safety group that we work with, but I do think more resources would be beneficial. There is already a lot more talent and money now than there was even four years ago. In percentage terms, there has been a rapid growth trajectory, even though in absolute terms it’s still a very small field.

马丁·福特：您认为超级智能问题应该更多地出现在公共领域吗？您想看到美国总统候选人谈论超级智能吗？

MARTIN FORD: Do you think that superintelligence concerns should be more in the public sphere? Do you want to see presidential candidates in the United States talking about superintelligence?

尼克·博斯特罗姆：不是。现在寻求州和政府的参与还为时过早，因为目前还不清楚人们希望他们做些什么才能在目前这个阶段有所帮助。首先需要澄清和更好地理解问题的性质，而且即使没有政府的介入，也可以做很多工作。我认为目前不需要对机器超级智能制定任何特定的规定。近期的人工智能应用涉及各种各样的事情，政府可能会在其中发挥各种作用。

NICK BOSTROM: Not really. It’s still a bit too early to seek involvement from states and governments because right now it’s not exactly clear what one would want them to do that would be helpful at this point in time. The nature of the problem first needs to be clarified and understood better, and there’s a lot of work that can be done without having governments come in. I don’t see any need right now for any particular regulations with respect to machine superintelligence. There are all kinds of things related to near-term AI applications where there might be various roles for governments to play.

如果你要在城市各处使用无人机，或者在街道上使用自动驾驶汽车，那么大概就需要有一个框架来规范它们。人工智能对经济和劳动力市场的影响程度也应该引起管理教育系统或制定经济政策的人的兴趣。我仍然认为超级智能有点超出了政客的职权范围，他们主要考虑的是任期内可能发生的事情。

If you’re going to have flying drones everywhere in the cities, or self-driving cars on the streets, then there presumably needs to be a framework that regulates them. The extent that AI will have an impact on the economy and the labor market is also something that should be of interest to people running education systems or setting economic policy. I still think superintelligence is a little bit outside the purview of politicians, who mainly think about what might happen during their tenure.

马丁·福特：那么，当埃隆·马斯克说超级智能比朝鲜的威胁更大时，这种言论是否会使情况变得更糟？

MARTIN FORD: So, when Elon Musk says superintelligence is a bigger threat than North Korea, could that rhetoric potentially make things worse?

尼克·博斯特罗姆：如果你过早介入此事，认为会出现大规模军备竞赛，从而导致竞争更加激烈，呼吁谨慎和全球合作的声音被边缘化，那么是的，这实际上可能会让事情变得更糟而不是更好。我认为，我们可以等到政府在超级智能方面真正需要和希望政府做的事情明确具体后，再尝试启动它们。在那之前，我们还有很多工作可以做，例如与人工智能开发社区以及从事人工智能的公司和学术机构合作，所以让我们暂时继续打好基础吧。

NICK BOSTROM: If you are getting into this prematurely, with a view to there being a big arms race, which could lead to a more competitive situation where voices for caution and global cooperation get sidelined, then yes, that could actually make things worse rather than better. I think one can wait until there is a clear concrete thing that one actually would need and want governments to do in relation to superintelligence, and then one can try to get them activated. Until that time, there’s still a huge amount of work that we can do, for example, in collaboration with the AI development community and with companies and academic institutions that are working with AI, so let’s get on with that groundwork for the time being.

马丁·福特：您是如何进入人工智能社区的？您是如何开始对人工智能产生兴趣的？您的职业生涯是如何发展到现在的？

MARTIN FORD: How did you come to your role in the AI community? How did you first become interested in AI, and how did your career develop to the point it’s at right now?

尼克·博斯特罗姆：从我记事起，我就一直对人工智能感兴趣。我在大学里学习人工智能，后来又学习了计算神经科学，还学习了理论物理等其他学科。我之所以这样做，是因为我认为，首先，人工智能技术最终可以改变世界；其次，从智力上来说，弄清楚大脑或计算机如何产生思维是非常有趣的。

NICK BOSTROM: I’ve been interested in artificial intelligence for as long as I can remember. I studied artificial intelligence, and later computational neuroscience, at university, as well as other topics, like theoretical physics. I did this because I thought that firstly, AI technology could eventually be transformative in the world, and secondly because it’s very interesting intellectually to try to figure out how thinking is produced by the brain or in a computer.

我在 20 世纪 90 年代中期发表了一些关于超级智能的著作，2006 年我有机会在牛津大学创建了人类未来研究所 (FHI)。我和同事们一起全职研究未来技术对人类未来的影响，特别关注机器智能的未来，有些人甚至可能称之为痴迷。这促使我在 2014 年出版了《超级智能：路径、危险、策略》一书。目前，我们在 FHI 内有两个小组。一个小组专注于对齐问题的技术计算机科学工作，因此试图设计可扩展控制方法的算法。另一个小组专注于治理、政策、道德和机器智能进步的社会影响。

I published some work about superintelligence in the mid-1990s, and I had the opportunity in 2006 to create the Future of Humanity Institute (FHI) at Oxford University. Together with my colleagues, I work full-time on the implications of future technologies for the future of humanity, with a particular focus—some might say an obsession—on the future of machine intelligence. That then resulted in 2014 in my book Superintelligence: Paths, Dangers, Strategies. Currently, we have two groups within the FHI. One group focuses on technical computer science work on the alignment problem, so trying to craft algorithms for scalable control methods. The other group focuses on governance, policy, ethics and the social implications of advances in machine intelligence.

马丁·福特：您在人类未来研究所的工作中关注的是各种生存风险，而不仅仅是与人工智能相关的危险，对吗？

MARTIN FORD: In your work at the Future of Humanity Institute you’ve focused on a variety of existential risks, not just AI-related dangers, right?

尼克·博斯特罗姆：没错，但是我们也在关注生存的机会，我们并没有忽视技术的优势。

NICK BOSTROM: That’s right, but we’re also looking at the existential opportunities, we are not blind to the upside of technology.

马丁·福特：请告诉我您所考虑过的其他一些风险，以及为什么您选择将重点放在机器智能上。

MARTIN FORD: Tell me about some of the other risks you’ve looked at, and why you’ve chosen to focus so much on machine intelligence above all.

尼克·博斯特罗姆：在 FHI，我们感兴趣的是真正宏观的问题，即那些可能从根本上改变人类状况的事物。我们并不是想研究明年的 iPhone 会是什么样子，而是研究那些可能改变人类某些基本参数的事物——这些问题决定了地球上起源的智能生命的未来命运。从这个角度来看，我们感兴趣的是生存风险——可能永久摧毁人类文明的事物——以及可能永久影响我们未来轨迹的事物。我认为技术可能是这种从根本上重塑人类的最合理来源，而在技术中，只有少数技术可能带来生存风险或生存机遇；人工智能可能是其中最重要的。FHI 还有一个小组研究生物技术带来的生物安全风险，我们更感兴趣的是如何将这些不同的考虑因素结合起来——我们称之为宏观战略。

NICK BOSTROM: At the FHI, we’re interested in really big-picture questions, the things that could fundamentally change the human condition in some way. We’re not trying to study what next year’s iPhone might be like, but instead things that could change some fundamental parameter of what it means to be human—questions that shape the future destiny of Earth-originating intelligent life. From that perspective, we are interested in existential risk—things that could permanently destroy human civilization—and also things that could permanently shape our trajectory into the future. I think technology is maybe the most plausible source for such fundamental reshapers of humanity, and within technology there are just a few that plausibly present either existential risks or existential opportunities; AI might be the foremost amongst those. FHI also has a group working on the biosecurity risks coming out of biotechnology, and we’re interested more generally in how you put these different considerations together—a macro strategy, as we call it.

为什么特别关注人工智能？我认为，如果人工智能能够成功实现其最初的目标，即一直以来的目标不仅仅是自动化特定任务，而是在机器中复制使人类变得聪明的通用学习能力和规划能力，那么这实际上就是人类需要创造的最后一项发明。如果实现，它将不仅在人工智能方面产生巨大影响，而且会对所有技术领域产生巨大影响，事实上，在所有人类智能目前有用的领域也是如此。

Why AI in particular? I think that if AI were to be successful at its original goal, which all along has been not just to automate specific tasks but to replicate in machine substrates the general-purpose learning ability and planning ability that makes us humans smart, then that would quite literally be the last invention that humans ever needed to make. If achieved, it would have enormous implications not just in AI, but across all technological fields, and indeed all areas where human intelligence currently is useful.

马丁·福特：比如说气候变化？这是你列出的生存威胁之一吗？

MARTIN FORD: What about climate change, for example? Is that on your list of existential threats?

尼克·博斯特罗姆：情况并非如此，部分原因是我们更愿意把重点放在我们认为我们的努力可能会产生重大影响的地方，而这些领域往往是问题相对被忽视的领域。目前，世界各地有许多人致力于应对气候变化。此外，很难想象地球温度升高几度会导致人类灭绝或永久摧毁未来。因此，出于这些和其他一些原因，这并不是我们努力的核心，尽管我们偶尔可能会通过试图总结人类面临的挑战的总体情况来侧面审视它。

NICK BOSTROM: Not so much, partly because we prefer to focus where we think our efforts might make a big difference, which tends to be areas where the questions have been relatively neglected. There are tons of people currently working on climate change across the world. Also, it’s hard to see how the planet getting a few degrees warmer would cause the extinction of the human species, or permanently destroy the future. So, for those and some other reasons, that’s not been at the center of our own efforts, although we might cast a sideways glance at it on occasion by trying to sum up the overall picture of the challenges that humanity confronts.

马丁·福特：所以，你会认为先进人工智能带来的风险实际上比气候变化更大，而我们在这些问题上的资源和投资分配不正确？这听起来像是一个非常有争议的观点。

MARTIN FORD: So, you would argue that the risk from advanced AI is actually more significant than from climate change, and that we’re allocating our resources and investment in these questions incorrectly? That sounds like a very controversial view.

尼克·博斯特罗姆：我确实认为存在一些分配不当的情况，而且这种现象不仅仅发生在这两个领域之间。总体而言，我认为人类文明并没有如此明智地分配我们的注意力。如果我们想象人类拥有一定数量的关注资本、关注筹码或恐惧筹码，我们可以将这些关注资本或恐惧筹码分散到威胁人类文明的各种事物上，我认为我们在选择分配这些关注筹码方面并没有那么老练。

NICK BOSTROM: I do think that there is some misallocation, and it’s not just between those two fields in particular. In general, I don’t think that we as a human civilization allocate our attention that wisely. If we imagine humans as having an amount of concern capital, chips of concern or fear that we can spread around on different things that threaten human civilization, I don’t think we are that sophisticated in how we choose to allocate those concern chips.

如果回顾上个世纪，你会发现在任何一个特定时间点，全球都会出现一个重大问题，所有受过良好教育的人都应该关注这个问题，而这个问题会随着时间的推移而发生变化。也许 100 年前是劣生论，知识分子担心人类种群的退化。然后在冷战期间，核战争显然是一个大问题，然后有一段时间是人口过剩。目前，我认为是全球变暖，尽管过去几年人工智能也悄然兴起。

If you look back over the last century, there has been at any given point in time maybe one big global concern that all intellectually educated people are supposed to be fixated on, and it’s changed over time. So maybe 100 years ago, it was dysgenics, where intellectuals were worrying about the deterioration of the human stock. Then during the Cold War, obviously nuclear Armageddon was a big concern, and then for a while, it was overpopulation. Currently, I would say it’s global warming, although AI has, over the last couple of years, been creeping up there.

马丁·福特：这可能很大程度上是因为伊隆·马斯克等人的谈论。你认为他如此直言不讳是件好事吗？还是说这有被过度炒作或吸引不知情的人参与讨论的危险？

MARTIN FORD: That’s perhaps largely due to the influence of people like Elon Musk talking about it. Do you think that’s a positive thing that he’s been so vocal, or is there a danger that it becomes overhyped or it draws uninformed people into the discussion?

尼克·博斯特罗姆：我认为到目前为止，人们对人工智能的评价都是积极的。当我写书的时候，我发现整个人工智能话题被忽视的程度令人震惊。有很多人在研究人工智能，但很少有人考虑如果人工智能成功了会发生什么。这也不是那种你可以和人们进行严肃交谈的话题，因为他们会把它当作科幻小说，但现在情况已经改变。

NICK BOSTROM: I think so far it has been met positively. When I was writing my book, it was striking how neglected the whole topic of AI was. There were a lot of people working on AI, but very few people thinking about what would happen if AI were to succeed. It also wasn’t the kind of topic you could have a serious conversation with people about because they would dismiss it as just science fiction, but that’s now changed.

我认为这很有价值，也许由于这已经成为一个更主流的话题，现在可以对诸如对齐问题之类的问题进行研究并发表技术论文。有许多研究小组正在这样做，包括 FHI，我们与 DeepMind 举行联合技术研究研讨会，OpenAI 也有许多人工智能安全研究人员，还有其他团体，如伯克利的机器智能研究所。我不确定，如果没有首先提高整个挑战的知名度，是否会有如此多的人才涌入这个领域。今天最需要的不是进一步的警惕或进一步的焦虑，人们大声呼吁关注，现在的挑战更多的是将这种现有的担忧和兴趣引导到建设性的方向，并继续开展工作。

I think that’s valuable, and maybe as a consequence of this having become a more mainstream topic, it’s now possible to do research and publish technical papers on things like the alignment problem. There are a number of research groups doing just that, including here at the FHI, where we have joint technical research seminars with DeepMind, also OpenAI has a number of AI safety researchers, and there are other groups like the Machine Intelligence Research Institute at Berkeley. I’m not sure whether there would have been as much talent flowing into this field unless the profile of the whole challenge had first been raised. What is most needed today is not further alarm or further hand-wringing with people screaming for attention, the challenge now is more to channel this existing concern and interest in constructive directions and to get on with the work.

马丁·福特：您担心的机器智能风险是否真的都取决于实现 AGI 以及更高级的超级智能？狭义人工智能的风险可能很大，但并不像您所说的那样危及生命。

MARTIN FORD: Is it true to say that the risks you worry about in terms of machine intelligence are really all dependent on achieving AGI and beyond that, superintelligence? The risks associated with narrow AI are probably significant, but not what you would characterize as existential.

尼克·博斯特罗姆：没错。我们也对这些近期的机器智能应用感兴趣，这些应用本身就很有趣，也值得讨论。我认为，当这两个不同的背景，即近期和长期被混为一谈时，麻烦就来了。

NICK BOSTROM: That’s correct. We do also have some interest in these more near-term applications of machine intelligence, which are interesting in their own right and also worth having a conversation about. I think the trouble arises when these two different contexts, the near term, and the long term get thrown into the same pot and confused.

马丁·福特：未来五年左右我们需要担心哪些近期风险？

MARTIN FORD: What are some of the near-term risks that we need to worry about over the next five years or so?

尼克·博斯特罗姆：在短期内，我认为主要有一些事情让我非常兴奋，并期待着推出。在短期内，好处远远大于坏处。只要看看经济和所有拥有更智能算法可以带来积极影响的领域。即使是一个低调、无聊的算法在大型物流中心的后台运行，更准确地预测需求曲线，也能让你减少库存，从而降低消费者的价格。

NICK BOSTROM: In the near term, I think primarily there are things that I would be very excited about and look forward to having roll out. In the near-term context, the upside far outweighs the downside. Just look across to the economy and at all the areas where having smarter algorithms could make a positive difference. Even a low-key, boring algorithm running in the background in a big logistic center predicting demand curves more accurately would enable you to reduce the amount of stock, and therefore cut prices for consumers.

在医疗保健领域，能够识别猫、狗和人脸的神经网络同样可以识别 X 射线图像中的肿瘤，并帮助放射科医生做出更准确的诊断。这些神经网络可能在后台运行，帮助优化患者流程并跟踪结果。几乎任何领域都可以，而且可能都有创造性的方法可以很好地利用机器学习中出现的这些新技术。

In healthcare, the same neural networks that can recognize cats, dogs, and faces could recognize tumors in x-ray images and assist radiologists in making more accurate diagnoses. Those neural networks might run in the background and help optimize patient flows and track outcomes. You could name almost any area, and there would probably be creative ways to use these new techniques that are emerging from machine learning to good effect.

我认为这是一个非常令人兴奋的领域，为企业家提供了许多机会。从科学的角度来看，开始了解智力是如何运作的以及大脑和这些神经系统是如何进行感知的，这确实令人兴奋。

I think that’s a very exciting field, with a lot of opportunity for entrepreneurs. From a scientific point of view as well, it’s really exciting to begin to understand a little bit about how intelligence works and how perception is performed by the brain and in these neural systems.

马丁·福特：很多人担心诸如自主武器之类的东西在近期会带来的风险，这些武器可以自行决定杀死谁。你支持禁止这类武器吗？

MARTIN FORD: A lot of people worry about the near-term risks of things like autonomous weapons that can make their own decisions about who to kill. Do you support a ban on weapons of those types?

尼克·博斯特罗姆：如果世界能够避免立即陷入另一场军备竞赛，那将是件好事，因为在军备竞赛中，人们会花费大量资金来完善杀手机器人。从广义上讲，我更希望机器智能用于和平目的，而不是开发新的毁灭人类的方法。我认为，如果我们仔细研究，就会发现人们到底希望条约禁止什么就不那么清楚了。

NICK BOSTROM: It would be positive if the world could avoid immediately jumping into another arms race, where huge amounts of money are spent perfecting killer robots. Broadly speaking, I’d prefer that machine intelligence is used for peaceful purposes, and not to develop new ways of destroying us. I think if one zooms in, it becomes a little bit less clear exactly what it is that one would want to see banned by a treaty.

有人主张人类必须参与其中，我们不应该让自主无人机自行做出目标决策，也许这是可能的。然而，另一种选择是，你拥有完全相同的系统，但不是无人机决定发射导弹，而是一个 19 岁的年轻人坐在弗吉尼亚州阿灵顿的电脑屏幕前，负责在屏幕上弹出“开火”窗口时按下红色按钮。如果这就是人类监督的意义所在，那么它与整个系统完全自主之间到底有多大区别就不清楚了。我认为也许更重要的是要有一定的责任，如果出了问题，你可以踢某人的屁股。

There’s a move to say that humans must be in the loop and that we should not have autonomous drones make targeting decisions on their own, and maybe that is possible. However, the alternative is that you have exactly the same system in place, but instead of the drone deciding to fire a missile, a19-year-old sits in Arlington, Virginia in front of a computer screen and has the job that whenever a window pops up on the screen saying “Fire,” they need to press a red button. If that’s what human oversight amounts to, then it’s not clear that it really makes that much of a difference from having the whole system be completely autonomous. I think maybe more important is that there is some accountability, and there’s somebody whose butt you can kick if things go wrong.

马丁·福特：你可以想象在某些情况下，自动机器可能更可取。例如，考虑到警务而非军事应用，我们在美国就曾发生过看似警察种族歧视的事件。在这种情况下，一个设计合理的人工智能驱动的机器人系统不会有偏见。它还会准备好先挡子弹，然后再开枪，而这对人类来说真的不是一个选择。

MARTIN FORD: There are certain situations you can imagine where an autonomous machine might be preferable. Thinking of policing rather than military applications, we’ve had incidents in the United States of what appears to be police racism, for example. A properly designed AI-driven robotic system in a situation like that would not be biased. It would also be prepared to take a bullet first, and shoot second, which is really not an option for a human being.

尼克·博斯特罗姆：我们最好不要在彼此之间发动战争，但如果要开战，也许最好是机器杀人，而不是年轻人向其他年轻人开枪。如果要打击特定的战斗人员，也许你可以进行精确打击，只杀死你想杀死的人，而不会对平民造成附带伤害。这就是为什么我说，当考虑到具体情况时，整体计算会变得更加复杂，考虑到人们希望在致命自主武器方面实施的具体规则或协议是什么。

NICK BOSTROM: Preferably we shouldn’t be fighting any wars between ourselves at all, but if there are going to be wars, maybe it’s better if it’s machines killing machines rather than young men shooting holes in other young men. If there are going to be strikes against specific combatants, maybe you can make precision strikes that only kill the people you’re trying to kill, and don’t create collateral damage with civilians. That’s why I’m saying that the overall calculation becomes a little bit more complex when one considers the specifics, and what exactly the rule or agreement is that one would want to be implemented with regard to lethal autonomous weapons.

其他应用领域也引发了有趣的伦理问题，例如监视或数据流管理、营销和广告，这些问题对于人类文明的长期结果可能与无人机用于杀戮或伤害人员的更直接应用同样重要。

There are other areas of application that also raise interesting ethical questions such as in surveillance, or the management of data flows, marketing, and advertising, which might matter as much for the long-term outcome of human civilization as these more direct applications of drones to kill or injure people.

马丁·福特：您认为对这些技术进行监管具有一定作用吗？

MARTIN FORD: Do you feel there is a role for regulation of these technologies?

尼克·博斯特罗姆：当然需要一些监管。如果你要拥有杀手无人机，你肯定不希望任何老罪犯能够使用装有面部识别软件的无人机在五公里外轻易暗杀政府官员。同样，你也不希望业余爱好者在机场上空驾驶无人机造成严重延误。我相信，随着越来越多的无人机穿越人类出于其他目的而旅行的空间，将需要某种军事框架。

NICK BOSTROM: Some regulation, for sure. If you’re going to have killer drones, you don’t want any old criminal to be able to easily assassinate public officials from five kilometers away using a drone with facial recognition software. Likewise, you don’t want to have amateurs flying drones across airports and causing big delays. I’m sure a form of military framework will be required as we get more of these drones traversing spaces where humans are traveling for other purposes.

马丁·福特：您的《超级智能：路径、危险、策略》一书出版已有四年了。事情进展是否如您所预期的那样顺利？

MARTIN FORD: It’s been about four years since your book Superintelligence: Paths, Dangers, Strategies was published. Are things progressing at the rate that you expected?

尼克·博斯特罗姆：过去几年的进展比预期的要快，尤其是深度学习取得了巨大进步。

NICK BOSTROM: Progress has been faster than expected over the last few years, with big advances in deep learning in particular.

马丁·福特：您在书中用一张表格来说明，计算机击败世界上最优秀的围棋选手还需要十年时间，所以那大约是 2024 年。但事实证明，这确实发生在您出版这本书的两年后。

MARTIN FORD: You had a table in your book where you said that having a computer beat the best Go player in the world was a decade out, so that would have been roughly 2024. As things turned out, it actually occurred just two years after you published the book.

尼克·博斯特罗姆：我认为我所说的是，如果进展继续保持过去几年的相同速度，那么人们可以预期在本书出版十年后，围棋冠军机器就会出现。然而，进展速度比这更快，部分原因是他们为解决围棋问题付出了专门的努力。DeepMind 接受了挑战，并指派了一些优秀人才来完成这项任务，并投入了大量的计算能力。但这无疑是一个里程碑，也展示了这些深度学习系统的惊人能力。

NICK BOSTROM: I think the statement I made was that if progress continued at the same rate as it had been going over the last several years, then one would expect a Go Grand Champion machine to occur about a decade after the book was written. However, the progress was faster than that, partly because there was a specific effort toward solving Go. DeepMind took on the challenge and assigned some good people to the task, and put a lot of computing power onto it. It was certainly a milestone, though, and a demonstration of the impressive capabilities of these deep learning systems.

马丁·福特：您认为我们与 AGI 之间的主要里程碑或障碍是什么？

MARTIN FORD: What are the major milestones or hurdles that you would point to that stand between us and AGI?

尼克·博斯特罗姆：机器学习仍面临几大挑战，比如需要更好的无监督学习技术。如果你想想成年人是如何了解我们所做的一切事情的，你会发现其中只有一小部分是通过明确的指导完成的。大部分都是我们观察正在发生的事情，并利用这种感官反馈来改进我们的世界模型。我们在幼儿时期也进行了大量试验和错误，把不同的东西撞在一起，看看会发生什么。

NICK BOSTROM: There are several big challenges remaining in machine learning, such as needing better techniques for unsupervised learning. If you think about how adult humans come to know all the things we do, only a small fraction of that is done through explicit instruction. Most of it is by us just observing what’s going on and using that sensory feed to improve our world models. We also do a lot of trial and error as toddlers, banging different things into one another and seeing what happens.

为了获得真正高效的机器智能系统，我们还需要能够更多地利用无监督和无标记数据的算法。作为人类，我们倾向于以因果关系来组织我们的大量世界知识，而目前的神经网络实际上并没有做到这一点。它更多的是在复杂模式中寻找统计规律，而不是真正将其组织成可以对其他对象产生各种因果影响的对象。所以，这将是一个方面。

In order to get really highly effective machine intelligent systems, we also need algorithms that can make more use of unsupervised and unlabeled data. As humans, we tend to organize a lot of our world knowledge in causal terms, and that’s something that is not really done much by current neural networks. It’s more about finding statistical regularities in complex patterns, but not really organizing that as objects that can have various kinds of causal impacts on other objects. So, that would be one aspect.

我还认为，规划和其他一些领域也需要进步，但并不是说没有关于如何实现这些目标的想法。现有的技术有限，在各个方面做得都相对较差，我认为，我们只需要在这些领域进行大量改进，就能实现人类通用智能。

I also think that there are advances needed in planning and a number of other areas as well, and it is not as if there are no ideas out there on how to achieve these things. There are limited techniques available that can do various aspects of these things relatively poorly, and I think that there just needs to be a great deal of improvement in those areas in order for us to get all the way to full human general intelligence.

马丁·福特：DeepMind 似乎是极少数专注于 AGI 的公司之一。您认为还有其他公司在做重要的工作，可能与 DeepMind 竞争吗？

MARTIN FORD: DeepMind seems to be one of the very few companies that’s focused specifically on AGI. Are there other players that you would point to that are doing important work, that you think may be competitive with what DeepMind is doing?

尼克·博斯特罗姆：DeepMind 无疑是其中的佼佼者，但还有很多地方正在开展机器学习方面的激动人心的研究，或者最终可能有助于实现通用人工智能。谷歌本身也拥有另一个世界级的人工智能研究团队，即谷歌大脑。其他大型科技公司现在都有自己的人工智能实验室：Facebook、百度和微软都在进行大量人工智能研究。

NICK BOSTROM: DeepMind is certainly among the leaders, but there are many places where there is exciting work being done on machine learning or work that might eventually contribute to achieving artificial general intelligence. Google itself has another world-class AI research group in the form of Google Brain. Other big tech companies now have their own AI labs: Facebook, Baidu, and Microsoft have quite a lot of research in AI going on.

在学术界，有很多优秀的地方。加拿大有蒙特利尔和多伦多，这两所大学都是世界领先的深度学习大学，而伯克利、牛津、斯坦福和卡内基梅隆等大学也有很多该领域的研究人员。这不仅仅是西方的事情，中国等国家也在大力投资建设国内能力。

In academia, there are a number of excellent places. Canada has Montreal and Toronto, both of which are world-leading deep learning universities, and the likes of Berkeley, Oxford, Stanford, and Carnegie Mellon also have a lot of researchers in the field. It’s not just a Western thing, countries like China are investing greatly in building up their domestic capacity.

马丁·福特：不过，这些并不是专门针对 AGI 的。

MARTIN FORD: Those are not focused specifically on AGI, though.

尼克·博斯特罗姆：是的，但界限很模糊。在目前公开致力于 AGI 的那些团体中，除了 DeepMind 之外，我想 OpenAI 是另一个值得关注的团体。

NICK BOSTROM: Yes, but it’s a fuzzy boundary. Among those groups currently overtly working towards AGI, aside from DeepMind, I guess OpenAI would be another group that one could point to.

马丁·福特：您认为图灵测试是确定我们是否已达到 AGI 的好方法吗，或者我们是否需要另一项智力测试？

MARTIN FORD: Do you think the Turing test is a good way to determine if we’ve reached AGI, or do we need another test for intelligence?

尼克·博斯特罗姆：如果你想要的是一个粗略的标准来衡量你何时完全成功，那还不算太糟。我说的是图灵测试的全面而困难的版本。你可以让专家对系统进行一个小时的询问，或类似的事情。我认为这是一个人工智能完全问题。除了开发通用人工智能外，没有其他方法可以解决这个问题。如果你感兴趣的是衡量进展速度，或者建立基准来了解你的人工智能研究团队下一步要努力的方向，那么图灵测试可能不是一个好的目标。

NICK BOSTROM: It’s not so bad if what you want is a rough-and-ready criterion for when you have fully succeeded. I’m talking about a full-blown, difficult version of the Turing test. Something where you can have experts interrogate the system for an hour, or something like that. I think that’s an AI-complete problem. It can’t be solved other than by developing general artificial intelligence. If what you’re interested in is gauging the rate of progress, say, or establishing benchmarks to know what to shoot for next with your AI research team, then the Turing test is maybe not such a good objective.

马丁·福特：因为如果规模较小的话它就会变成一种噱头吗？

MARTIN FORD: Because it turns into a gimmick if it’s at a smaller scale?

尼克·博斯特罗姆：是的。有一种方法可以做到正确，但太难了，我们现在根本不知道该怎么做。如果你想在图灵测试上取得渐进式进展，你得到的将是这些系统，它们插入了大量预设答案、巧妙的技巧和花招，但实际上并没有让你更接近真正的通用人工智能。如果你想在实验室中取得进展，或者如果你想衡量世界进步的速度，那么你需要其他基准，这些基准更能让我们走得更远，最终将带来完全通用的人工智能。

NICK BOSTROM: Yes. There’s a way of doing it right, but that’s too difficult, and we don’t know at all how to do that right now. If you wanted incremental progress on the Turing test, what you would get would be these systems that have a lot of canned answers plugged in, and clever tricks and gimmicks, but that actually don’t move you any closer to real AGI. If you want to make progress in the lab, or if you want to measure the rate of progress in the world, then you need other benchmarks that plug more into what is actually getting us further down the road, and that will eventually lead to fully general AI.

马丁·福特：那么意识呢？它是智能系统自动产生的吗？还是一种完全独立的现象？

MARTIN FORD: What about consciousness? Is that something that might automatically emerge from an intelligent system, or is that an entirely independent phenomenon?

尼克·博斯特罗姆：这取决于你对意识的定义。意识的一个含义是拥有一种功能性自我意识的能力，也就是说，你能够将自己塑造成世界上的演员，并反思不同事物如何改变你作为一个主体。你可以认为自己会随着时间而持续存在。这些事情或多或少是创建更智能系统的副作用，这些系统可以构建更好的模型来描述现实的各个方面，包括它们自己。

NICK BOSTROM: It depends on what you mean by consciousness. One sense of the word is the ability to have a functional form of self-awareness, that is, you’re able to model yourself as an actor in the world and reflect on how different things might change you as an agent. You can think of yourself as persisting through time. These things come more or less as a side effect of creating more intelligent systems that can build better models of all kinds of aspects of reality, and that includes themselves.

“意识”一词的另一个含义是我们认为具有道德意义的现象体验场。例如，如果某人实际上有意识地遭受痛苦，那么这在道德上就是一件坏事。它的含义不仅仅是他们倾向于逃避有害刺激，因为他们实际上在内心将其体验为一种主观感受。很难知道这种现象体验是否会作为使机器系统变得更智能的副作用而自动出现。甚至可能设计出没有感质但仍非常强大的机器系统。鉴于我们实际上并不十分清楚道德相关意识形式的必要和充分条件，我们必须接受机器智能可能获得意识的可能性，甚至可能早在它们达到人类水平或超级智能之前。

Another sense of the word “consciousness” is this phenomenal experiential field that we have that we think has moral significance. For example, if somebody is actually consciously suffering, then it’s a morally bad thing. It means something more than just that they tend to run away from noxious stimuli because they actually experience it inside of themselves as a subjective feeling. It’s harder to know whether that phenomenal experience will automatically arise just as a side effect of making machine systems smarter. It might even be possible to design machine systems that don’t have qualia but could still be very capable. Given that we don’t really have a very clear grasp of what the necessary and sufficient conditions are for morally relevant forms of consciousness, we must accept the possibility that machine intelligences could attain consciousness, maybe even long before they become human-level or superintelligent.

我们认为许多非人类动物拥有更多相关形式的经验。即使是像老鼠这样简单的动物，如果你想对老鼠进行医学研究，也必须遵循一套协议和指导方针。例如，在对老鼠进行手术之前，你必须对它进行麻醉，因为我们认为如果不麻醉就把它切开，它会很痛苦。如果我们拥有机器智能系统，比如说，它的行为和认知复杂性与老鼠相同，那么似乎有一个问题：到那时，它是否也会开始达到意识水平，从而赋予它一定程度的道德地位，并限制我们对它的所作所为。至少我们似乎不应该断然否定这种可能性。单是它可能有意识这一可能性可能已经足以让我们承担一些义务，至少如果这些义务很容易做到的话，这些义务将使机器拥有更高质量的生活。

We think many non-human animals have more of the relevant forms of experience. Even with something as simple as a mouse, if you want to conduct medical research on mice, there is a set of protocols and guidelines that you have to follow. You have to anesthetize a mouse before you perform surgery on it, for example, because we think it would suffer if you just carved it up without anesthesia. If we have machine-intelligent systems, say, with the same behavioral repertoire and cognitive complexity as a mouse, then it seems to be a live question whether at that point it might not also start to reach levels of consciousness that would give it some degree of moral status and limit what we can do to it. At least it seems we shouldn’t be dismissing that possibility out of hand. The mere possibility that it could be conscious might already be sufficient grounds for some obligations on our part to do, at least if they’re easy to do, things that will make the machine have a better-quality life.

马丁·福特：所以，从某种意义上说，这里的风险是双向的？我们担心人工智能会伤害我们，但也有风险，我们可能会奴役一个有意识的实体或让它受苦。在我看来，我们永远没有明确的方法来知道机器是否真的有意识。没有什么比图灵测试更能测试意识了。我相信你有意识，因为你和我一样是同一个物种，我相信我有意识，但你和机器没有那种联系。这是一个很难回答的问题。

MARTIN FORD: So, in a sense, the risks here run both ways? We worry about the risk of AI harming us, but there’s also the risk that perhaps we’re going to enslave a conscious entity or cause it to suffer. It sounds to me that there is no definitive way that we’re ever going to know if a machine is truly conscious. There’s nothing like the Turing test for consciousness. I believe you’re conscious because you’re the same species I am, and I believe I’m conscious, but you don’t have that kind of connection with a machine. It’s a very difficult question to answer.

尼克·博斯特罗姆：是的，我认为这很难。我不会说物种成员身份是我们用来假设意识的主要标准，有很多人类没有意识。也许他们处于昏迷状态，或者是胎儿，或者他们可能脑死亡，或者处于深度麻醉状态。大多数人还认为你可以是非人类，例如，某些动物，比如说，有不同程度和形式的意识体验。因此，我们能够将其投射到我们自己物种之外，但我认为，如果数字思维真的存在，那么将必要的道德考量扩展到数字思维对人类同理心来说将是一个挑战。

NICK BOSTROM: Yes, I think it is difficult. I wouldn’t say species membership is the main criterion here that we use to posit consciousness, there are a lot of human beings that are not conscious. Maybe they are in a coma, or they are fetuses, or they could be brain dead, or under deep anesthesia. Most people also think you can be a non-human being, for instance, certain animals, let us say, have various degrees and forms of conscious experience. So, we are able to project it outside our own species, but I think it is true that it will be a challenge for human empathy to extend the requisite level of moral consideration to digital minds, should such come to exist.

我们对待动物已经够难的了。我们对待动物的方式，尤其是在肉类生产方面，有很多不足之处，而且动物有脸，还会吱吱叫！如果微处理器内部有一个看不见的过程，人类就很难认识到其中可能存在值得考虑的有意识的思想。即使在今天，这似乎也是那些你无法认真对待的疯狂话题之一。它就像是哲学研讨会上的讨论，而不是真正的问题，就像算法歧视或杀手无人机一样。

We have a hard enough time with animals. Our treatment of animals, particularly in meat production, leaves much to be desired, and animals have faces and can squeak! If you have an invisible process inside a microprocessor, it’s going to be much harder for humans to recognize that there could be a sentient mind in there that deserves consideration. Even today, it seems like one of those crazy topics that you can’t really take seriously. It’s like a discussion for a philosophical seminar rather than a real issue, like algorithmic discrimination is, or killer drones.

最终，它需要脱离只有专业哲学家才会谈论的疯狂话题，成为可以进行合理公开辩论的话题。这需要逐步实现，但我认为也许是时候开始影响这种转变了，就像过去几年人工智能可能对人类状况产生什么影响的话题已经从科幻小说变成了更主流的话题一样。

Ultimately, it needs to be moved out of this sphere of crazy topics that only professional philosophers talk about, and into a topic that you could have a reasonable public debate about. It needs to happen gradually, but I think maybe it’s time to start affecting that shift, just as the topic of what AI might do for the human condition has moved from science-fiction into a more mainstream conversation over the last few years.

马丁·福特：您认为人工智能对就业市场和经济可能产生什么影响？您认为人工智能可能带来多大的颠覆？您是否认为人工智能是我们需要高度关注的问题？

MARTIN FORD: What do you think about the impact on the job market and the economy that artificial intelligence might have? How big a disruption do you think that could be and do you think that’s something we need to be giving a lot of attention to?

尼克·博斯特罗姆：我认为，在短期内，人们可能会倾向于夸大人工智能对劳动力市场的影响。真正大规模地推出人工智能系统并产生重大影响还需要时间。但随着时间的推移，我确实认为机器学习的进步将对人类劳动力市场产生越来越大的影响，如果人工智能完全成功，那么人工智能基本上可以做任何事情。在某些方面，人工智能的最终目标是实现完全失业。我们之所以从事技术开发和自动化，是因为我们不必付出太多努力就能实现既定的结果。你可以用更少的资源做更多的事情，这就是技术的本质。

NICK BOSTROM: In the very short term, I think that there might be a tendency to exaggerate the impacts on the labor market. It is going to take time to really roll out systems on a large enough scale to have a big impact. Over time, though, I do think that advances in machine learning will have an increasingly large impact on human labor markets and if you fully succeed with artificial intelligence, then yes, artificial intelligence could basically do everything. In some respects, the ultimate goal is full unemployment. The reason why we do technology, and why we do automation is so that we don’t have to put in so much effort to achieve a given outcome. You can do more with less, and that’s the gestalt of technology.

马丁·福特：这是乌托邦式的愿景。那么，您是否支持以基本收入作为确保每个人都能享受所有这些进步成果的机制？

MARTIN FORD: That’s the utopian vision. So, would you support, for example, a basic income as a mechanism to make sure that everyone can enjoy the fruits of all this progress?

尼克·博斯特罗姆：随着时间的推移，一些类似功能可能会变得越来越受欢迎。如果人工智能真的成功了，我们解决了技术控制问题，并进行了合理的治理，那么就会出现巨大的经济爆炸式增长。即使是其中的一小部分也足以让每个人过上真正美好的生活，所以似乎至少应该这样做。如果我们开发出超级智能，无论我们喜欢与否，我们都会承担这一发展带来的部分风险。因此，如果事情进展顺利，每个人都应该获得一些好处，这似乎是公平的。

NICK BOSTROM: Some functional analog of that could start to look increasingly desirable over time. If AI truly succeeds, and we resolve the technical control problem and have some reasonable governance, then an enormous bonanza of explosive economic growth takes place. Even a small slice of that would be ample enough to give everybody a really great life, so it seems one should at the minimum do that. If we develop superintelligence, we will all carry a slice of the risk of this development, whether we like it or not. It seems only fair, then, that everybody should also get some slice of the upside if things go well.

我认为这应该是机器超级智能在世界上的应用愿景的一部分；至少其中很大一部分应该是为了全人类的共同利益。这也与对开发者的私人激励一致，但如果我们真的中了大奖，那么蛋糕将非常大，我们应该确保每个人都拥有极好的生活质量。这可以采取某种全民基本收入的形式，也可以采取其他方案，但最终结果应该是每个人都在经济资源方面获得巨大收益。超级智能还可以带来其他好处，比如更好的技术、更好的医疗保健等等。

I think that should be part of the vision of how machine superintelligence should be used in the world; at least a big chunk of it should be for the common good of all of humanity. That’s also consistent with having private incentives for developers, but the pie, if we really hit the jackpot, would be so large that we should make sure that everybody has a fantastic quality of life. That could take the form of some kind of universal basic income or there could be other schemes, but the net result of that should be that everybody sees a great gain in terms of their economic resources. There will also be other benefits—like better technologies, better healthcare, and so forth—that superintelligence could enable.

马丁·福特：有人担心中国会率先或与我们同时实现 AGI，对此您怎么看？在我看来，无论哪个文化开发了这项技术，其价值观都很重要。

MARTIN FORD: What are your thoughts on the concern that China could reach AGI first, or at the same time as us? It seems to me that the values of whatever culture develops this technology do matter.

尼克·博斯特罗姆：我认为，哪种文化最先发展出这种安全感可能并不重要。发展这种安全感的人或群体的能力如何，以及他们是否有机会小心谨慎，才是更重要的。这是赛车运动中令人担忧的问题之一，因为有很多不同的竞争对手竞相争抢着要第一个到达终点线——在激烈的比赛中，你不得不不顾一切。谁在安全方面投入的精力最少，谁就能赢得比赛，这将是非常不理想的情况。

NICK BOSTROM: I think it might matter less which particular culture happens to develop it first. It matters more how competent the particular people or group that are developing it are, and whether they have the opportunity to be careful. This is one of the concerns with a racing dynamic, where you have a lot of different competitors racing to get to some kind of finish line first—in a tight race you are forced to throw caution to the wind. The race would go to whoever squanders the least effort on safety, and that would be a very undesirable situation.

我们宁愿让开发第一个超级智能的人在开发过程结束时有选择暂停六个月，甚至几年的时间来重新检查他们的系统并安装他们能想到的任何额外保护措施。只有这样，他们才会慢慢地、谨慎地将系统的能力提升到超人的水平。你不希望他们因为某个竞争对手紧随其后而仓促行事。当思考未来超级智能出现时人类最理想的战略形势是什么时，似乎一个重要的要求是尽可能缓和竞争态势。

We would rather have whoever it is that develops the first superintelligence to have the option at the end of the development process to pause for six months, or maybe a couple of years to double-check their systems and install whatever extra safeguards they can think of. Only then would they slowly and cautiously amplify the system’s capabilities up to the superhuman level. You don’t want them to be rushed by the fact that some competitor is nipping at their heels. When thinking about what the most desirable strategic situation for humanity is when superintelligence arises in the future, it seems that one important desideratum is that the competitive dynamics should be allayed as much as possible.

马丁·福特：如果我们真的有一个“快速起飞”的场景，让智能可以不断自我改进，那么先发优势就非常大。谁先到达那里，谁就可能根本无法追赶，所以，你所说的那种竞争，对谁来说，存在着巨大的激励，这不是一件好事。

MARTIN FORD: If we do have a “fast takeoff” scenario where the intelligence can recursively improve itself, though, then there is an enormous first-mover advantage. Whoever gets there first could essentially be uncatchable, so there’s a huge incentive for exactly the kind of competition that you’re saying isn’t a good thing.

尼克·博斯特罗姆：在某些情况下，确实可以有这样的动态，但我认为我之前提到的以可​​信的承诺追求这一目标并将其用于全球利益的观点在这里很重要，这不仅从道德角度，而且从降低比赛动态强度的角度来看也很重要。如果所有参赛者都觉得即使他们没有赢得比赛，他们仍然会受益匪浅，那就太好了。这样一来，最终做出某种安排就更加可行，让领先者可以在不着急的情况下获得胜利。

NICK BOSTROM: In certain scenarios, yes, you could have dynamics like that, but I think the earlier point I made about pursuing this with a credible commitment to using it for the global good is important here, not only from an ethical point of view but also from the point of view of reducing the intensity of the racing dynamic. It would be good if all the competitors feel that even if they don’t win the race, they’re still going to benefit tremendously. That will then make it more feasible to have some arrangement in the end where the leader can get a clean shot at this without being rushed.

马丁·福特：这需要某种形式的国际协调，而人类在这方面的记录并不那么好。与化学武器禁令和核不扩散法案相比，听起来人工智能在核实人们没有作弊方面将是一个更大的挑战，即使你确实有某种协议。

MARTIN FORD: That calls for some sort of international coordination, and humanity’s track record isn’t that great. Compared to the chemical weapons ban and the nuclear non-proliferation act, it sounds like AI would be an even greater challenge in terms of verifying that people aren’t cheating, even if you did have some sort of agreement.

尼克·博斯特罗姆：在某些方面，这会更具挑战性，而在另一些方面，挑战性可能较小。人类的游戏经常围绕稀缺性展开——资源非常有限，如果一个人或一个国家拥有这些资源，那么其他人就没有这些资源。有了人工智能，在许多方面就有了丰富的机会，这可以更容易地形成合作安排。

NICK BOSTROM: In some respects it would be more challenging, and in other respects maybe less challenging. The human game has often been played around scarcity—there is a very limited set of resources, and if one person or country has those resources, then somebody else does not have them. With AI there is the opportunity for abundance in many respects, and that can make it easier to form cooperative arrangements.

马丁·福特：您认为我们会解决这些问题吗？而且人工智能总体上会成为一股积极的力量吗？

MARTIN FORD: Do you think that we will solve these problems and that AI will be a positive force overall?

尼克·博斯特罗姆：我既充满希望，又充满恐惧。我想强调一下它的好处，无论是短期还是长期。由于我的工作和我的书，人们总是问我风险和弊端，但我内心深处也非常兴奋，渴望看到这项技术可以实现的所有有益用途，我希望这能给世界带来巨大的福祉。

NICK BOSTROM: I’m full of both hopes and fears. I would like to emphasize the upsides here, both in the short term and longer term. Because of my job and my book, people always ask me about the risks and downsides, but a big part of me is also hugely excited and eager to see all the beneficial uses that this technology could be put to and I hope that this could be a great blessing for the world.

尼克·博斯特罗姆 是牛津大学的教授，也是人类未来研究所的创始主任。他还负责人工智能治理项目。尼克曾在哥德堡大学、斯德哥尔摩大学和伦敦国王学院学习，后于 2000 年获得伦敦政治经济学院哲学博士学位。他著有约 200 部著作，包括《人因偏见》（2002 年）、《全球灾难性风险》（2008 年）、《人类增强》（2009 年）和《超级智能：路径、危险和策略》（2014 年），后者是《纽约时报》畅销书。

NICK BOSTROM is a Professor at Oxford University, where he is the founding Director of the Future of Humanity Institute. He also directs the Governance of Artificial Intelligence Program. Nick studied at the University of Gothenburg, Stockholm University and Kings College London prior to receiving his PhD in philosophy from the London School of Economics in 2000. He is the author of some 200 publications, including Anthropic Bias (2002), Global Catastrophic Risks (2008), Human Enhancement (2009), and Superintelligence: Paths, Dangers, Strategies (2014), a New York Times bestseller.

尼克拥有物理学、人工智能、数理逻辑和哲学背景。他是尤金·R·甘农奖（每年从哲学、数学、艺术和其他人文学科以及自然科学领域选出一人）获得者。他曾两次入选《外交政策》杂志的全球 100 位思想家名单；他被列入《展望》杂志的世界思想家名单，是所有领域前 15 名中最年轻的人，也是排名最高的分析哲学家。他的著作已被翻译成 24 种语言。他的作品有 100 多种翻译和再版。

Nick has a background in physics, artificial intelligence, and mathematical logic as well as philosophy. He is recipient of a Eugene R. Gannon Award (one person selected annually worldwide from the fields of philosophy, mathematics, the arts and other humanities, and the natural sciences). He has been listed on Foreign Policy’s Top 100 Global Thinkers list twice; and he was included on Prospect magazine’s World Thinkers list, the youngest person in the top 15 from all fields and the highest-ranked analytic philosopher. His writings have been translated into 24 languages. There have been more than 100 translations and reprints of his works.

Facebook 副总裁兼首席人工智能科学家、纽约大学计算机科学教授

VP & CHIEF AI SCIENTIST, FACEBOOK PROFESSOR OF COMPUTER SCIENCE, NYU

三十多年来，Yann LeCun 一直致力于人工智能和机器学习的学术和工业研究。在加入 Facebook 之前，Yann 曾在 AT&T 的贝尔实验室工作，在那里他因开发卷积神经网络而备受赞誉，这是一种受大脑视觉皮层启发的机器学习架构。Yann 与 Geoff Hinton 和 Yoshua Bengio 一起，是少数几位研究人员中的一员，他们的努力和坚持直接导致了深度学习神经网络的革命。

Yann LeCun has been involved in the academic and industry side of AI and Machine Learning for over 30 years. Prior to joining Facebook, Yann worked at AT&T’s Bell Labs, where he is credited with developing convolutional neural networks—a machine learning architecture inspired by the brain’s visual cortex. Along with Geoff Hinton and Yoshua Bengio, Yann is part of a small group of researchers whose effort and persistence led directly to the current revolution in deep learning neural networks.

马丁·福特：让我们直接开始讨论过去十年左右发生的深度学习革命。这场革命是如何开始的？我认为它是神经网络技术的一些改进、计算机速度的提高和可用训练数据量的激增的综合产物，对吗？

MARTIN FORD: Let’s jump right in and talk about the deep learning revolution that’s been unfolding over the past decade or so. How did that get started? Am I right that it was the confluence of some refinements to neural network technology, together with much faster computers and an explosion in the amount of training data available?

YANN LECUN：是的，但比这更刻意。随着 1986-87 年反向传播算法的出现，人们能够训练多层神经网络，这是旧模型无法做到的。这引发了一股兴趣浪潮，一直持续到 1995 年左右，然后逐渐消退。

YANN LECUN: Yes, but it was more deliberate than that. With the emergence of the backpropagation algorithm in 1986-87, people were able to train neural nets with multiple layers, which was something that the old models didn’t do. This resulted in a wave of interest that lasted right through to around 1995 before petering out.

然后在 2003 年，杰弗里·辛顿 (Geoffrey Hinton)、约书亚·本吉奥 (Yoshua Bengio) 和我聚在一起说，我们知道这些技术最终会胜出，我们需要聚在一起制定一个计划来重新引起社区对这些方法的兴趣。这就是深度学习的起源。如果你愿意的话，这是一个蓄意的阴谋。

Then in 2003, Geoffrey Hinton, Yoshua Bengio, and I got together and said, we know these techniques are eventually going to win out, and we need to get together and hash out a plan to renew the community interest in these methods. That’s what became deep learning. It was a deliberate conspiracy, if you will.

马丁·福特：回首过去，你曾想象过自己会取得多大的成功吗？如今，人们认为人工智能和深度学习是同义词。

MARTIN FORD: Looking back, did you imagine the extent to which you would be successful? Today, people think artificial intelligence and deep learning are synonymous.

YANN LECUN：是也不是。是的，因为我们知道这些技术最终会在计算机视觉、语音识别以及其他一些领域占据主导地位。但我们没有意识到它会成为深度学习的代名词。

YANN LECUN: Yes and no. Yes, in the sense that we knew eventually those techniques would come to the fore for computer vision, speech recognition, and maybe a couple of other things—but no, we didn’t realize it would become synonymous with deep learning.

我们没有意识到，整个行业对它的兴趣会如此之大，以至于会催生出一个全新的行业。我们没有意识到，公众对它的兴趣会如此之大，它不仅会彻底改变计算机视觉和语音识别，还会改变自然语言理解、机器人技术、医学成像分析，并且会让自动驾驶汽车真正发挥作用。这确实让我们感到惊讶。

We didn’t realize that there would be so much of an interest from the wider industry that it would create a new industry altogether. We didn’t realize that there would be so much interest from the public, and that it would not just revolutionize computer vision and speech recognition, but also natural language understanding, robotics, medical imaging analysis, and that it would enable self-driving cars that actually work. That took us by surprise, that’s for sure.

早在 90 年代初，我就认为这种进步会稍早一些但更为渐进地发生，而不是像 2013 年左右那样发生大革命。

Back in the early ‘90s, I would have thought that that this kind of progress would have happened slightly earlier but more progressively, rather than the big revolution that occurred around 2013.

马丁·福特：您是如何开始对人工智能和机器学习感兴趣的？

MARTIN FORD: How did you first become interested in AI and machine learning?

YANN LECUN：小时候，我对科学、工程和一些重大科学问题很感兴趣，比如生命、智能、人类的起源。人工智能让我着迷，尽管在 20 世纪 60 年代和 70 年代，人工智能在法国还不存在。尽管我对这些问题很感兴趣，但高中毕业后，我相信自己最终会成为一名工程师，而不是科学家，所以我开始学习工程学。

YANN LECUN: As a kid, I was interested in science and engineering and the big scientific questions—life, intelligence, the origin of humanity. Artificial intelligence was something that fascinated me, even though it didn’t really exist as a field in France during the 1960s and 1970s. Even with a fascination for those questions, when I finished high school I believed that I would eventually become an engineer rather than a scientist, so I began my studies in the field of engineering.

在我学习的早期，大约在 1980 年，我偶然发现了一本哲学书，它是发展心理学家让·皮亚杰和语言学家诺姆·乔姆斯基之间辩论的记录，名为《语言与学习：让·皮亚杰和诺姆·乔姆斯基之间的辩论》。这本书包含了一场非常有趣的辩论，辩论涉及先天和后天的概念以及语言和智力的出现。

Early on in my studies, around 1980, I stumbled on a philosophy book which was a transcription of a debate between Jean Piaget, the developmental psychologist, and Noam Chomsky, the linguist, called, Language and Learning: The Debate Between Jean Piaget and Noam Chomsky. The book contained a really interesting debate between the concepts of nature and nurture and the emergence of language and intelligence.

在这场辩论中，站在皮亚杰一边的是西摩·帕普特，他是麻省理工学院计算机科学教授，参与了早期机器学习的研究，可以说，他在 20 世纪 60 年代末的第一波神经网络浪潮中彻底消灭了机器学习领域。10 年后，他却对一种非常简单的机器学习模型——感知器大加赞赏，这种模型是在 20 世纪 50 年代发明的，他在 20 世纪 60 年代一直在研究它。那是我第一次读到学习机器的概念，我对机器学习的想法非常着迷。我认为学习是智能不可或缺的一部分。

On the side of Piaget in the debate was Seymour Papert, who was a professor at MIT in computer science and who was involved with early machine learning and arguably actually killed the field off in the first wave of neural nets in the late 1960s. Here he was, 10 years later, singing the praise of a very simple machine learning model called the perceptron that had been invented in the 1950s, and that he had been working on in the 1960s. That was the first time I read about the concept of a learning machine, and I was absolutely fascinated by the idea that a machine could learn. I thought learning was an integral part of intelligence.

读本科时，我查阅了所有能找到的关于机器学习的文献，并做了几个项目。我发现西方没有人研究神经网络。一些日本研究人员正在研究后来被称为神经网络的东西，但西方没有人研究，因为这个领域在 60 年代末被著名的美国人工智能研究员 Seymour Papert 和 Marvin Minsky 扼杀了。

As an undergrad, I dug up all the literature I could find about machine learning and did a couple of projects on it. I discovered that nobody in the West was working on neural nets. A few Japanese researchers were working on what became known as neural networks, but no one in the West was, because the field had been killed in the late ‘60s in part by Seymour Papert and Marvin Minsky, the famous American AI researcher.

我继续独自研究神经网络，并于 1987 年获得博士学位，论文题目为“ Modeles connexionnistes de l'apprentissage ” （联结主义学习模型）。我的导师 Maurice Milgram 实际上并没有研究这个主题，他直接告诉我，“我可以成为你的正式导师，但我无法在技术上帮助你。”

I carried on working on neural nets on my own, and I did a PhD in 1987 titled, Modeles connexionnistes de l’apprentissage (Connectionist learning models). My advisor, Maurice Milgram, was not actually working on this topic, and he told me outright, “I can be your official advisor, but I can’t help you technically.”

我在工作中发现，在 20 世纪 80 年代初期，世界各地有一个研究神经网络的群体，我与他们建立了联系，并最终与 David Rumelhart、Geoffrey Hinton 等人一起发现了诸如反向传播之类的东西。

I discovered through my work that in the early 1980s, there was a community of people around the world who were working on neural nets, and I connected with them and ended up discovering things like backpropagation in parallel with people like David Rumelhart and Geoffrey Hinton.

马丁·福特：那么，20 世纪 80 年代初期，加拿大在这个领域进行了大量研究？

MARTIN FORD: So, in the early 1980s there was a lot of research in this area going on in Canada?

YANN LECUN：不，那是在美国。加拿大当时还没有开展此类研究。20 世纪 80 年代初，Geoffrey Hinton 是加州大学圣地亚哥分校的博士后，与 David Rumelhart 和 Jay McClelland 等认知科学家一起工作。最终，他们出版了一本书，用简单的神经网络和计算模型解释心理学。Geoffrey 随后成为卡内基梅隆大学的副教授，直到 1987 年才搬到多伦多。那时我也搬到了多伦多，在他的实验室做了一年的博士后。

YANN LECUN: No, this was the United States. Canada was not on the map for this type of research yet. In the early 1980s, Geoffrey Hinton was a postdoc at the University of California, San Diego where he was working with cognitive scientists like David Rumelhart and Jay McClelland. Eventually they published a book explaining psychology by simple neural nets and models of computation. Geoffrey then became Associate Professor at Carnegie Mellon University, and only moved to Toronto in 1987. That’s when I also moved to Toronto, where I was a postdoc in his lab for one year.

马丁·福特：20 世纪 80 年代初，我还是一名计算机工程专业的本科生，我根本不记得自己接触过神经网络。这是一个当时就存在的概念，但当时它绝对被边缘化了。现在，到了 2018 年，情况发生了巨大变化。

MARTIN FORD: I was an undergraduate studying computer engineering in the early 1980s, and I don’t recall much exposure to neural networks at all. It was a concept that was out there, but it was definitely very much marginalized. Now, in 2018, that has changed dramatically.

YANN LECUN：这比被边缘化还要糟糕。在 70 年代和 80 年代初期，神经网络在社区中是被唾弃的。你甚至不能发表一篇提到“神经网络”这个词的论文，因为它会立即被你的同行拒绝。

YANN LECUN: It was worse than marginalized. In the ‘70s and early ‘80s it was anathema within the community. You couldn’t publish a paper that even mentioned the phrase neural networks because it would immediately be rejected by your peers.

事实上，Geoffrey Hinton 和 Terry Sejnowski 在 1983 年发表了一篇非常著名的论文，名为《最佳感知推理》，该论文描述了一种早期的深度学习或神经网络模型。Hinton 和 Sejnowski 不得不使用暗语来避免提及这是一个神经网络。甚至他们的论文标题都是隐晦的；一切都很奇怪！

In fact, Geoffrey Hinton and Terry Sejnowski published a very famous paper in 1983 called, Optimal Perceptual Inference, which described an early deep learning or neural network model. Hinton and Sejnowski had to use code words to avoid mentioning that it was a neural network. Even the title of their paper was cryptic; it was all very strange!

马丁·福特：您最知名的主要创新之一是卷积神经网络。您能解释一下它是什么吗？它与深度学习中的其他方法有何不同？

MARTIN FORD: One of the main innovations you’re known for is the convolutional neural network. Could you explain what that is and how it’s different from other approaches in deep learning?

YANN LECUN：卷积神经网络的动机是构建一个适合识别图像的神经网络。事实证明，它对语音识别和语言翻译等多种任务都很有用。它在某种程度上受到了动物或人类视觉皮层结构的启发。

YANN LECUN: The motivation for convolutional neural networks was building a neural network that was appropriate for recognizing images. It turned out to be useful for a wide-range of tasks, such as speech recognition and language translation. It’s somewhat inspired by the architecture of the visual cortex in animals or humans.

戴维·休伯尔 (David Hubel) 和托尔斯滕·维塞尔 (Torsten Wiesel) 在 20 世纪 50 年代和 60 年代在神经科学领域做出了一些诺贝尔奖研究，研究内容涉及视觉皮层中的神经元所发挥的功能类型以及它们之间的相互联系。

David Hubel and Torsten Wiesel did some Nobel prize-winning work in neuroscience in the 1950s and 1960s about the type of functions that the neurons in the visual cortex perform and how they’re connected with each other.

卷积网络是一种将神经元相互连接的特殊方式，这样处理起来就适合图像之类的事物。我应该补充一点，我们通常不称它们为神经元，因为它们并不是生物神经元的准确反映。

A convolutional network is a particular way of connecting the neurons with each other in such a way that the processing that takes place is appropriate for things like images. I should add that we don’t normally call them neurons because they’re not really an accurate reflection of biological neurons.

神经元连接的基本原理是，它们被组织成多层，第一层中的每个神经元都与输入图像中的一小块像素相连。每个神经元计算其输入的加权和。权重是通过学习修改的量。神经元只能看到输入的一小部分像素，并且有一大堆神经元会查看同一个小窗口。然后，有一大堆神经元会查看另一个略微偏移的窗口，但这一堆神经元执行的操作与另一堆相同。如果你有一个神经元在一个窗口中检测特定的主题，那么你将有另一个神经元在下一个窗口中检测完全相同的主题，并且其他神经元会查看整个图像的所有窗口。

The basic principle of how the neurons are connected is that they’re organized in multiple layers and each neuron in the first layer is connected with a small patch of pixels in the input image. Each neuron computes a weighted sum of its inputs. The weights are the quantities that are modified by learning. The neurons only see a tiny window of pixels of the input, and there’s a whole bunch of neurons that look at the same little window. Then, there’s a whole bunch of neurons that look at another slightly shifted window, but this bunch performs the same operation as the other bunch. If you have a neuron that detects a particular motif in one window, you’re going to have another neuron that detects exactly the same motif in the next window and other neurons for all windows across the image.

一旦将所有这些神经元放在一起，并且您就会意识到它们进行什么样的数学运算，这种运算就称为离散卷积，这就是为什么它被称为卷积网络。

Once you put all those neurons together and you realize what kind of mathematical operation they do, that operation is called a discrete convolution, which is why this is called a convolutional net.

这是第一层，然后是第二层，它是非线性层 - 基本上是一个阈值，如果卷积层计算的加权和高于或低于阈值，每个神经元就会打开或关闭。

That’s the first layer, and then there’s a second layer, which is a non-linearity layer—basically a threshold where each neuron turns on or turns off if the weighted sum computed by the convolution layer is above or below the threshold.

最后，还有第三层，它执行所谓的池化操作。我不会详细介绍它，但它基本上起到确保当输入图像略微移动或变形时，输出响应不会发生太大变化的作用。这是一种对输入图像中对象的扭曲移动或变形建立一点不变性的方法。

Finally, there’s a third layer that performs what’s called a pooling operation. I’m not going to cover it in detail, but it basically plays a role in making sure that when the input image is slightly shifted or deformed, the output responses don’t change that much. That’s a way of building a bit of invariance to distortion shifts or deformation of the object in the input image.

卷积网络基本上就是这种类型的层的堆叠——卷积、非线性、池化。堆叠多层，当到达顶层时，你就有了应该检测单个物体的神经元。

The convolutional net is basically a stack of layers of this type—convolution, non-linearity, pooling. You stack multiple layers of those, and by the time you get to the top, you have neurons that are supposed to detect individual objects.

如果你在图像中放入一匹马的图像，你可能会有一个神经元被打开，然后你会有一个用于识别汽车、人、椅子和所有其他你可能想要识别的类别的神经元。

You might have a neuron that turns on if you put an image of a horse in the image, and then you have one for cars, people, chairs, and all other categories you might want to recognize.

诀窍在于，这个神经网络的功能是由神经元之间的连接强度和权重决定的，而这些都不是编程出来的，而是经过训练的。

The trick is that the function that this neural network is doing is determined by the strength of the connections between the neurons, the weights, and those are not programmed; they’re trained.

这就是你在训练神经网络时学到的东西。你给它看一匹马的图像，如果它没有说“马”，你就告诉它这是错的，这是它应该说的答案。然后通过使用反向传播算法，它会调整网络中所有连接的所有权重，这样下次你展示同一匹马的图像时，输出就会更接近你想要的，你会对数千张图像重复这一操作。

This is what is learned when you train the neural net. You show it the image of a horse, and if it doesn’t say “horse,” you tell it that it’s wrong and here is the answer that it should have said. Then by using the backpropagation algorithm, it adjusts all the weights of all the connections in the network so that next time you show the same image of a horse, the output would be closer to the one you want, and you keep doing this for thousands of images.

马丁·福特：通过给网络提供猫或马等图像来训练网络的过程就是所谓的监督学习，对吗？可以说监督学习是当今的主流方法，并且需要大量数据吗？

MARTIN FORD: That process of training a network by giving it images of cats or horses, and so on, is what’s called supervised learning, correct? Is it true to say that supervised learning is the dominant approach today, and that it takes huge amounts of data?

YANN LECUN：没错。如今几乎所有深度学习应用都使用监督学习。

YANN LECUN: Exactly. Almost all of the applications of deep learning today use supervised learning.

监督学习是指在训练机器时，你给它正确的答案，然后它自己修正并给出正确的答案。它的神奇之处在于，在训练之后，它大多数时候都会在训练过的类别中给出正确的答案，即使是它从未见过的图像也是如此。你说得对，这通常需要大量样本，至少在你第一次训练网络时是这样。

Supervised learning is when you give the correct answer to the machine when you’re training it, and then it corrects itself to give the correct answer. The magic of it is that after it’s been trained, it produces a correct answer most of the time in categories that it’s been trained on, even for images it’s never seen before. You’re correct, that does typically require a lot of samples, at least the first time you train the network.

马丁·福特：您如何看待该领域未来的发展？监督学习与人类儿童的学习方式截然不同。您可以指着一只猫说“那是一只猫”，这个样本可能足以让孩子学习。这与当今的人工智能截然不同。

MARTIN FORD: How do you see the field moving forward in the future? Supervised learning is very different from the way a human child learns. You could point at a cat once and say, “there’s a cat,” and that one sample might be enough for a child to learn. That’s dramatically different from where AI is today.

YANN LECUN：嗯，既是也不是。正如我所说，第一次训练卷积网络时，你会用数千张甚至数百万张不同类别的图像来训练它。如果你想添加一个新类别，例如，如果机器从未见过猫，而你想训练它识别猫，那么它只需要一些猫的样本。这是因为它已经被训练过识别任何类型的图像，并且知道如何表示图像；它知道什么是物体，并且它对各种物体了解很多。因此，要训练它识别新物体，你只需向它展示一些样本，并且只需训练几个顶层。

YANN LECUN: Well, yes and no. As I said, the first time you train a convolutional network you train it with thousands, possibly even millions of images of various categories. If you then want to add a new category, for example if the machine has never seen a cat and you want to train it to recognize cats, then it only requires a few samples of cats. That is because it has already been trained to recognize images of any type and it knows how to represent images; it knows what an object is, and it knows a lot of things about various objects. So, to train it to recognize a new object, you just show it a few samples, and you just need to train a couple of the top layers.

马丁·福特：那么，如果你训练一个网络来识别其他种类的动物，比如狗和熊，那么是否只需要少量的数据就可以识别猫？这似乎与孩子可能做的事情没什么不同。

MARTIN FORD: So, if you trained a network to recognize other kinds of animals like dogs and bears, then would it only take a small amount of data to get to a cat? That seems not so different from what a child is probably doing.

YANN LECUN：但情况有所不同，这是不幸的。儿童（以及动物）的学习方式是，他们大部分的学习都是在你告诉他们“这是一只猫”之前进行的。在生命的最初几个月里，婴儿通过观察学到了很多东西，而没有任何语言概念。他们通过观察和与世界的少量互动，学到了大量关于世界如何运作的知识。

YANN LECUN: But it is different, and that’s the unfortunate thing. The way a child learns (and animals, for that matter) is that most of the learning they do is before you can tell them, “this is a cat.” In the first few months of life, babies learn a huge amount by observation without having any notion of language. They learn an enormous amount of knowledge about how the world works just by observation and with a little interaction with the world.

这种对世界背景知识的大量积累是我们不知道如何用机器实现的。我们不知道该怎么称呼它，有些人称之为无监督学习，但这是一个含义丰富的术语。它有时被称为预测学习或推断学习。我称之为自我监督学习。这是一种你不为某项任务进行训练的学习，你只是观察世界并弄清楚它是如何运作的。

This sort of accumulation of enormous amounts of background knowledge about the world is what we don’t know how to do with machines. We don’t know what to call this, some people call this unsupervised learning, but it’s a loaded term. It’s sometimes called predictive learning, or imputative learning. I call it self-supervised learning. It’s the kind of learning where you don’t train for a task, you just observe the world and figure out how it works, essentially.

马丁·福特：强化学习，或者通过实践进行学习并对成功给予奖励，属于无监督学习的范畴吗？

MARTIN FORD: Would reinforcement learning, or learning by practice with a reward for succeeding, be in the category of unsupervised learning?

YANN LECUN：不，那是完全不同的类别。本质上有三个类别；它更像是一个连续体，但有强化学习、监督学习和自我监督学习。

YANN LECUN: No, that’s a different category altogether. There are three categories essentially; it’s more of a continuum, but there is reinforcement learning, supervised learning, and self-supervised learning.

强化学习是通过反复试验进行学习，成功时获得奖励，失败时则不会获得奖励。这种学习形式在最纯粹的形式下，在样本方面效率极低，因此在游戏中效果很好，因为你可以随意尝试多次，但在很多现实世界场景中却行不通。

Reinforcement learning is learning by trial and error, getting rewards when you succeed and not getting rewards when you don’t succeed. That form of learning in its purest form is incredibly inefficient in terms of samples, and as a consequence works well for games, where you can try things as many times as you want, but doesn’t work in many real-world scenarios.

你可以使用强化学习来训练机器下围棋或国际象棋。这种方法效果很好，例如 AlphaGo，但它需要大量的样本或试验。一台机器要达到良好的表现，基本上要比过去 3000 年全人类玩的游戏都多，如果你能做到这一点，效果会很好，但在现实世界中，这往往是不切实际的。

You can use reinforcement learning to train a machine to play Go or chess. That works really well, as we’ve seen with AlphaGo, for example, but it requires a ridiculous number of samples or trials. A machine has to basically play more games than all of humanity in the last 3,000 years to reach good performance, and it works really well if you can do that, but it is often impractical in the real world.

如果你想用强化学习来训练机器人抓取物体，那将需要花费大量的时间。人类可以在 15 个小时的训练中学会驾驶汽车，而且不会撞到任何东西。如果你想用目前的强化学习方法训练汽车自动驾驶，那么机器必须驾驶汽车冲下悬崖 10,000 次才能弄清楚如何避免这种情况。

If you want to use reinforcement learning to train a robot to grab objects, it will take a ridiculous amount of time to achieve that. A human can learn to drive a car in 15 hours of training without crashing into anything. If you want to use the current reinforcement learning methods to train a car to drive itself, the machine will have to drive off cliffs 10,000 times before it figures out how not to do that.

马丁·福特：我想这就是模拟的论点。

MARTIN FORD: I guess that’s the argument for simulation.

YANN LECUN：我不同意。这可能是支持模拟的论点，但也是支持这样一个事实的论点：我们人类能够进行的学习与纯粹的强化学习非常不同。

YANN LECUN: I don’t agree. It might be an argument for simulation, but it’s also an argument for the fact that the kind of learning that we can do as humans is very, very different from pure reinforcement learning.

它更类似于人们所说的基于模型的强化学习。在这种技术中，你有一个内部世界模型，可以让你预测当你朝某个方向转动方向盘时，汽车就会朝某个方向行驶，如果前面有另一辆车，你会撞上它，或者如果有悬崖，你会从悬崖上掉下去。你有这个预测模型，可以让你提前预测你的行为的后果。因此，你可以提前计划，避免采取导致不良后果的行动。

It’s more akin to what people call model-based reinforcement learning. This is where you have your internal model of the world that allows you to predict that when you turn the wheel in a particular direction then the car is going to go in a particular direction, and if another car comes in front you’re going to hit it, or if there is a cliff you are going to fall off that cliff. You have this predictive model that allows you to predict in advance the consequence of your actions. As a result, you can plan ahead and not take the actions that result in bad outcomes.

在这种情况下学习驾驶被称为基于模型的强化学习，这是我们真正不知道如何做的事情之一。它有一个名字，但没有真正的方法让它可靠地工作！大多数学习不是在强化中，而是在以自我监督的方式学习预测模型，这是我们今天不知道如何解决的主要问题。

Learning to drive in this context is called model-based reinforcement learning, and that’s one of the things we don’t really know how to do. There is a name for it, but there’s no real way to make it work reliably! Most of the learning is not in the reinforcement, it’s in learning the predictive models in a self-supervised manner, and that’s the main problem we don’t know how to solve today.

马丁·福特：这是你在 Facebook 工作时重点关注的领域吗？

MARTIN FORD: Is this an area that you’re focused on with your work at Facebook?

YANN LECUN：是的，这是我们在 Facebook 正在研究的事情之一。我们正在研究很多不同的事情，包括让机器通过观察不同的数据源来学习——学习世界是如何运转的。我们正在建立一个世界模型，这样也许某种形式的常识就会出现，也许这个模型可以用作一种预测模型，让机器能够像人类一样学习，而不必在成功之前尝试 10,000 次并失败。

YANN LECUN: Yes, it is one of the things that we’re working on at Facebook. We’re working on a lot of different things, including getting machines to learn by observation from different data sources—learning how the world works. We’re building a model of the world so that perhaps some form of common sense will emerge and perhaps that model could be used as kind of a predictive model that would allow a machine to learn the way people do without having to try and fail 10,000 times before they’ve succeeded.

马丁·福特：有些人认为，单靠深度学习是不够的，或者网络需要更多的结构，从一开始就需要某种智能设计。你似乎坚信智能会从相对通用的神经网络中自然而然地涌现出来。

MARTIN FORD: Some people argue that deep learning alone is not going to be enough, or that there needs to be more structure in the networks, some kind of intelligent design from the onset. You seem to be a strong believer in the idea that intelligence will emerge organically from relatively generic neural networks.

YANN LECUN：我觉得这样说有点夸张。每个人都同意需要某种结构，问题是需要多少结构以及需要什么样的结构。我想，当你说有些人认为应该有逻辑和推理之类的结构时，你可能指的是 Gary Marcus 和 Oren Etzioni。

YANN LECUN: I think that would be an exaggeration. Everybody agrees that there is a need for some structure, the question is how much, and what kind of structure is needed. I guess when you say that some people believe that there should be structures such as logic and reasoning, you’re probably referring to Gary Marcus and maybe Oren Etzioni.

今天早些时候，我与 Gary Marcus 就此进行了辩论。Gary 的观点在社区中并不被广泛接受，因为他一直在批判性地撰写有关深度学习的文章，但并没有为深度学习做出贡献。Oren Etzioni 的情况并非如此，因为他已经在这个领域工作了一段时间，但他的观点比 Gary 温和得多。不过，我们所有人都同意的一点是，需要某种结构。

I actually had a debate with Gary Marcus on this earlier today. Gary’s view isn’t particularly well accepted in the community because he’s been writing critically about deep learning, but he’s not been contributing to it. That’s not the case for Oren Etzioni because he’s been in the field for a while, but his view is considerably milder than Gary’s. The one thing all of us agree on, though, is that there is a need for some structure.

事实上，卷积网络的理念就是在神经网络中建立结构。卷积网络不是一张白纸，它们确实有一些结构。问题是，如果我们想要人工智能出现，而且我们谈论的是通用智能或人类级别的人工智能，我们需要多少结构？这就是人们的观点可能不同的地方，比如我们是否需要明确的结构来让机器操纵符号，或者我们是否需要明确的结构来表示语言中的层次结构。

In fact, the very idea of convolutional networks is to put a structure in neural networks. Convolutional networks are not a blank slate, they do have a little bit of structure. The question is, if we want AI to emerge, and we’re talking general intelligence or human-level AI, how much structure do we need? That’s where people’s views may differ, like whether we need explicit structures that will allow a machine to manipulate symbols, or if we need explicit structures for representing hierarchical structures in language.

我的许多同事，比如 Geoffrey Hinton 和 Yoshua Bengio，都同意从长远来看，我们不需要为此建立精确的特定结构。这可能在短期内有用，因为我们可能还没有找到一种用于自我监督学习的通用学习方法。因此，一种偷工减料的方法就是硬连线架构；这是一件非常好的事情。但从长远来看，我们并不清楚我们需要多少这样的架构。无论是视觉皮层还是前额叶皮层，皮层的微结构似乎都非常非常统一。

A lot of my colleagues, like Geoffrey Hinton and Yoshua Bengio, agree that in the long run we won’t need precise specific structures for this. It might be useful in the short term because we may not have figured out a general learning method for self-supervised learning. So, one way to cut corners is to hardwire the architecture; that is a perfectly fine thing to do. In the long run, though, it’s not clear how much of that we need. The microstructure of the cortex seems to be very, very uniform all over, whether you’re looking at the visual or prefrontal cortex.

马丁·福特：大脑是否使用类似反向传播的东西？

MARTIN FORD: Does the brain use something like backpropagation?

YANN LECUN：我们真的不知道。不过，还有比这更根本的问题。人们想出的大多数学习算法本质上都是最小化某个目标函数。

YANN LECUN: We don’t really know. There are more fundamental questions than that, though. Most of the learning algorithms that people have come up with essentially consist of minimizing some objective function.

我们甚至不知道大脑是否会最小化目标函数。如果大脑确实最小化了目标函数，那么它是否通过基于梯度的方法来做到这一点？大脑是否有某种方法来估计在哪个方向上修改其所有突触连接以改进这个目标函数？我们不知道。如果它估计了那个梯度，它是否通过某种形式的反向传播来实现？

We don’t even know if the brain minimizes an objective function. If the brain does minimize an objective function, does it do it through a gradient-based method? Does the brain have some way of estimating in which direction to modify all of its synaptic connections in such a way as to improve this objective function? We don’t know that. If it estimates that gradient, does it do it by some form of backpropagation?

它可能不是我们所知的反向传播，但它可能是梯度估计的一种近似形式，与反向传播非常相似。Yoshua Bengio 一直在研究生物学上可行的梯度估计形式，因此大脑对某些目标函数进行某种梯度估计并非完全不可能，我们只是不知道而已。

It’s probably not backpropagation as we know it, but it could be a form of approximation of gradient estimation that is very similar to backpropagation. Yoshua Bengio has been working on biologically plausible forms of gradient estimation, so it’s not entirely impossible that the brain does some sort of gradient estimation of some objective function, we just simply don’t know.

马丁·福特：您在 Facebook 还在研究哪些其他重要课题？

MARTIN FORD: What other important topics are you working on at Facebook?

YANN LECUN：我们正在进行大量机器学习的基础研究和问题研究，因此这些研究与应用数学和优化关系更密切。我们正在研究强化学习，我们也在研究一种称为生成模型的东西，这是一种自我监督或预测学习的形式。

YANN LECUN: We’re working on a lot of fundamental research and questions on machine learning, so things that have more to do with applied mathematics and optimization. We are working on reinforcement learning, and we are also working on something called generative models, which are a form of self-supervised or predictive learning.

马丁·福特：Facebook 是否正在致力于构建可以真正进行对话的系统？

MARTIN FORD: Is Facebook working on building systems that can actually carry out a conversation?

YANN LECUN：到目前为止，我提到的都是研究的基本主题，但还有很多应用领域。

YANN LECUN: What I’ve mentioned so far are the fundamental topics of research, but there are a whole bunch of application areas.

Facebook 在计算机视觉领域非常活跃，我认为我们可以声称拥有世界上最好的计算机视觉研究小组。这是一个成熟的小组，里面有很多很酷的活动。我们在自然语言处理方面投入了大量精力，包括翻译、摘要、文本分类——弄清楚文本谈论的主题，以及对话系统。实际上，对话系统是虚拟助手、问答系统等非常重要的研究领域。

Facebook is very active in computer vision, and I think we can claim to have the best computer vision research group in the world. It’s a mature group and there are a lot of really cool activities there. We’re putting quite a lot of work into natural language processing, and that includes translation, summarization, text categorization—figuring out what topic a text talks about, as well as dialog systems. Actually, dialog systems are a very important area of research for virtual assistants, question and answering systems, and so on.

马丁·福特：您是否预计有一天人工智能能够通过图灵测试？

MARTIN FORD: Do you anticipate the creation of an AI that someday could pass the Turing test?

YANN LECUN：这终究会发生，但图灵测试其实并不是一个有趣的测试。事实上，我认为目前人工智能领域的许多人并不认为图灵测试是一个好的测试。图灵测试太容易被欺骗了，而且在某种程度上，图灵测试已经过去了。

YANN LECUN: It’s going to happen at some point, but the Turing test is not actually an interesting test. In fact, I don’t think a lot of people in the AI field at the moment consider the Turing test to be a good test. It’s too easy to trick it, and to some extent, the Turing test has already been and gone.

作为人类，我们非常重视语言，因为我们习惯于通过语言与其他人讨论智能话题。然而，语言是智能的一种附带现象，当我这样说时，从事自然语言处理工作的同事们强烈反对！

We give a lot of importance to language as humans because we are used to discussing intelligent topics with other humans through language. However, language is sort of an epiphenomenon of intelligence, and when I say this, my colleagues who work on natural language processing disagree vehemently!

看看猩猩，它们本质上和我们一样聪明。它们拥有大量的常识和非常好的世界模型，它们可以像人类一样制造工具。然而，它们没有语言，它们不是群居动物，除了非语言的母子互动外，它们几乎不与其他物种成员互动。智力中有一整套与语言无关的成分，如果我们将人工智能简化为仅满足图灵测试，我们就忽略了这一点。

Look at orangutans, who are essentially almost as smart as we are. They have a huge amount of common sense and very good models of the world, and they can build tools, just like humans. However, they don’t have language, they’re not social animals, and they barely interact with other members of the species outside the non-linguistic mother-and-child interaction. There is a whole component of intelligence that has nothing to do with language, and we are ignoring this if we reduce AI to just satisfying the Turing test.

马丁·福特：通用人工智能的道路是什么？我们需要克服什么才能实现通用人工智能？

MARTIN FORD: What is the path to artificial general intelligence and what do we need to overcome to get there?

YANN LECUN：可能还有一些我们目前没有发现但最终会遇到的问题，但我认为我们需要弄清楚的一件事是婴儿和动物在生命最初几天、几周和几个月内通过观察了解世界如何运作的能力。

YANN LECUN: There are probably other problems that we do not see at the moment that we’re going to eventually encounter, but one thing I think we’ll need to figure out is the ability that babies and animals have to learn how the world works by observation in the first few days, weeks, and months of life.

在这段时间里，你会了解到世界是三维的。你会了解到，当你移动头部时，有些物体会以不同的方式移动到其他物体的前面。你会了解物体的永久性，因此你会了解到，当一个物体隐藏在另一个物体后面时，它仍然存在。随着时间的推移，你会了解重力、惯性和刚度——这些都是非常基本的概念，主要通过观察来学习。

In that time, you learn that the world is three-dimensional. You learn that there are objects that move in front of others in different ways when you move your head. You learn object permanence, so you learn that when an object is hidden behind another one, it’s still there. As time goes on, you learn about gravity, inertia, and rigidity—very basic concepts that are learnt essentially by observation.

婴儿没有太多的手段来对世界采取行动，但他们观察很多，并通过观察学到很多东西。幼年动物也是如此。它们可能有更多的固定内容，但非常相似。

Babies don’t have a huge amount of means to act on the world, but they observe a lot, and they learn a huge amount by observing. Baby animals also do this. They probably have more hardwired stuff, but it’s very similar.

在我们弄清楚如何进行这种无监督/自我监督/预测学习之前，我们不会取得重大进展，因为我认为这是学习足够多的世界背景知识的关键，这样常识才会出现。这是主要的障碍。这其中还有更多的技术子问题我无法深入探讨，比如不确定性下的预测，但这是主要问题。

Until we figure out how to do this unsupervised/self-supervised/predictive learning, we’re not going to make significant progress because I think that’s the key to learning enough background knowledge about the world so that common sense will emerge. That’s the main hurdle. There are more technical subproblems of this that I can’t get into, like prediction under uncertainty, but that’s the main thing.

我们还要花多长时间才能找到一种方法来训练机器，让它们通过观看 YouTube 视频来了解世界是如何运转的？目前还不清楚。我们可能在两年内取得突破，但可能还需要 10 年才能真正实现，也可能需要 10 年或 20 年。我不知道什么时候会发生，但我知道这一定会发生。

How long is it going to take before we figure out a way to train machines so that they learn how the world works by watching YouTube videos? That’s not entirely clear. We could have a breakthrough in two years that might take another 10 years to actually make it work, or it might take 10 or 20 years. I have no idea when it will happen, but I do know it has to happen.

这只是我们要翻越的第一座大山，我们不知道这座山后面还有多少座大山。可能还有其他重大问题和重大疑问我们还没有看到，因为我们还没有到达那里，那里是未知领域。

That’s just the first mountain we have to climb, and we don’t know how many mountains are behind it. There might be other huge issues and major questions that we do not see yet because we haven’t been there yet and it’s unexplored territory.

我们可能需要 10 年时间才能找到这种突破，并让它在现实世界产生影响，而这必须在我们达到人类水平的通用人工智能之前实现。问题是，一旦我们跨过这个障碍，还会出现什么其他问题？

It will probably take 10 years before we find this kind of breakthrough and before it has some consequence in the real world, and that has to happen way before we reach human-level artificial general intelligence. The question is, once we clear this hurdle, what other problems are going to pop up?

我们需要在这些系统中构建多少预先结构，才能使它们真正正常工作并保持稳定，并使它们具有内在动机，以便在人类周围表现得当？肯定会出现很多问题，所以 AGI 可能需要 50 年，也可能需要 100 年，我不太确定。

How much prior structure do we need to build into those systems for them to actually work appropriately and be stable, and for them to have intrinsic motivations so that they behave properly around humans? There’s a whole lot of problems that will absolutely pop up, so AGI might take 50 years, it might take 100 years, I’m not too sure.

马丁·福特：但您认为这是可以实现的吗？

MARTIN FORD: But you think it’s achievable?

YANN LECUN：哦，当然了。

YANN LECUN: Oh, definitely.

马丁·福特：您认为这是不可避免的吗？

MARTIN FORD: Do you think it’s inevitable?

YANN LECUN：是的，毫无疑问。

YANN LECUN: Yes, there’s no question about that.

马丁·福特：当你想到 AGI 时，它是有意识的，还是可能是完全没有意识经验的僵尸？

MARTIN FORD: When you think of an AGI, would it be conscious, or could it be a zombie with no conscious experience at all?

YANN LECUN：我们不知道这意味着什么。我们不知道意识是什么。我认为这不是问题。这是那些最终当你意识到事物实际上是如何运作时，你就会意识到这个问题并不重要的问题之一。

YANN LECUN: We don’t know what that means. We have no idea what consciousness is. I think it’s a non-problem. It’s one of those questions that in the end, when you realize how things actually work, you realize that question was immaterial.

早在 17 世纪，当人们发现眼球后部视网膜上的图像是颠倒的时，他们就对我们看到的是正反面的事实感到困惑。当你理解了之后需要什么样的处理，以及像素出现的顺序并不重要时，你就会意识到这是一个有趣的问题，因为它没有任何意义。这里也是一样。我认为意识是一种主观体验，它可能是聪明的一个非常简单的附带现象。

Back in the 17th century when people figured out that the image in the back of the eye on the retina forms upside down, they were puzzled by the fact that we see right-side up. When you understand what kind of processing is required after this, and that it doesn’t really matter in which order the pixels come, you realize it’s kind of a funny question because it doesn’t make any sense. It’s the same thing here. I think consciousness is a subjective experience and it could be a very simple epiphenomenon of being smart.

关于这种意识错觉的成因，有几种假说——因为我认为这是一种错觉。一种可能性是，我们的前额叶皮层中基本上有一个引擎，让我们能够模拟世界，而有意识地决定关注特定情况会根据当前情况配置该世界模型。

There are several hypotheses for what causes this illusion of consciousness—because I think it is an illusion. One possibility is that we have essentially a single engine in our prefrontal cortex that allows us to model the world, and a conscious decision to pay attention to a particular situation configures that model of the world for the situation at hand.

意识状态可以说是注意力的一种重要形式。如果我们的大脑比现在大十倍，而且我们没有一个引擎来模拟世界，而是有一大堆引擎，那么我们可能就不会有同样的意识体验。

The conscious state is sort of an important form of attention, if you will. We may not have the same conscious experience if our brain were ten times the size and we didn’t have a single engine to model the world, but a whole bunch of them.

马丁·福特：我们来谈谈人工智能带来的一些风险。您是否认为我们正处于一场重大经济动荡的边缘，并有可能造成大范围失业？

MARTIN FORD: Let’s talk about some of the risks associated with AI. Do you believe that we’re on the cusp of a big economic disruption with the potential for wide spread job losses?

YANN LECUN：我不是经济学家，但我显然也对这些问题感兴趣。我和很多经济学家谈过，也参加过许多会议，很多著名的经济学家都在讨论这些问题。首先，他们说人工智能是一种通用技术，简称 GPT。这意味着它是一种将渗透到经济各个角落的技术，几乎可以改变我们做所有事情的方式。这不是我说的，是他们说的。如果我这么说，我会显得自私或傲慢，除非我从其他知道他们在说什么的人那里听到过，否则我不会重复这句话。所以，他们是这么说的，在听到他们这么说之前，我并没有真正意识到是这样。他们说这是电力、蒸汽机或电动机规模的东西。

YANN LECUN: I’m not an economist, but I’m obviously interested in those questions, too. I’ve talked to a bunch of economists, and I’ve attended a number of conferences with a whole bunch of very famous economists who were discussing those very questions. First of all, what they say is that AI is what they call a general-purpose technology or GPT for short. What that means is that it’s a piece of technology that will diffuse into all corners of the economy and transform pretty much how we do everything. I’m not saying this; they are saying this. If I was saying this, I would sound self-serving or arrogant, and I would not repeat it unless I had heard it from other people who know what they’re talking about. So, they’re saying this, and I didn’t really realize that this was the case before I heard them say it. They say this is something on the scale of electricity, the steam engine, or the electric motor.

在与经济学家交谈之前，我担心的一件事是技术性失业问题。技术进步迅速，新经济所需的技能与人口的技能不匹配。整个人口比例突然间不具备合适的技能，他们被抛在后面。

One thing I’m worried about, and this was before talking to the economists, is the problem of technological unemployment. The idea that technology progresses rapidly and the skills that are required by the new economy are not matched by the skills of the population. A whole proportion of the population suddenly doesn’t have the right skills, and it’s left behind.

你可能会认为，随着技术进步的加速，会有更多的人被抛在后面，但经济学家们却认为，一项技术在经济中传播的速度实际上受到未接受过使用该技术培训的人数比例的限制。换句话说，被抛在后面的人越多，该技术在经济中传播的速度就越慢。这很有趣，因为这意味着邪恶有一种自我调节机制。除非有相当一部分人接受过真正利用人工智能技术的培训，否则我们不会广泛传播人工智能技术，他们用来证明这一点的例子就是计算机技术。

You would think that as technological progress accelerates, there’d be more and more people left behind, but what the economists say is that the speed at which a piece of technology disseminates in the economy is actually limited by the proportion of people who are not trained to use it. In other words, the more people are left behind, the less quickly the technology can diffuse in the economy. It’s interesting because it means that the evil has kind of a self-regulating mechanism in it. We’re not going to have widely disseminated AI technology unless a significant proportion of the population is trained to actually take advantage of it, and the example they use to demonstrate this is computer technology.

计算机技术在 20 世纪 60 年代和 70 年代兴起，但直到 20 世纪 90 年代才对经济生产力产生影响，因为人们花了很长时间才熟悉键盘、鼠标等，而软件和计算机也花了很长时间才变得足够便宜，从而具有大众的吸引力。

Computer technology popped up in the 1960s and 1970s but did not have an impact on productivity on the economy until the 1990s because it took that long for people to get familiar with keyboards, mice, etc., and for software and computers to become cheap enough for them to have mass appeal.

马丁·福特：我认为存在一个问题，即这一次是否与历史上的情况有所不同，因为机器现在具备了认知能力。

MARTIN FORD: I think there is a question of whether this time is different relative to those historical cases, because machines are taking on cognitive capability now.

现在，机器可以学习做很多常规、可预测的事情，而且我们劳动力中很大一部分从事的是可预测的工作。所以，我认为这次的颠覆可能比我们过去看到的更大。

You now have machines that can learn to do a lot of routine, predictable things, and a significant percentage of our workforce is engaged in things that are predictable. So, I think the disruption could turn out to be bigger this time than what we’ve seen in the past.

YANN LECUN：我其实并不这么认为。我不认为我们会因为这项技术的出现而面临大规模失业。我认为经济格局肯定会发生巨大变化，就像 100 年前大多数人口都在田里工作，而现在只有 2% 的人口在田里工作。

YANN LECUN: I don’t actually think that’s the case. I don’t think that we’re going to face mass unemployment because of the appearance of this technology. I think certainly the economic landscape is going to be vastly different in the same way that 100 years ago most of the population were working in the fields, and now it’s 2% of the population.

当然，在未来几十年里，你会看到这种转变，人们将不得不为此重新接受培训。我们需要某种形式的持续学习，这对每个人来说都不容易。不过，我不相信我们会失业。我听到一位经济学家说：“我们不会失业，因为我们不会没有问题。”

Certainly, over the next several decades, you’re going to see this kind of shift and people are going to have to retrain for it. We’ll need some form of continuous learning, and it’s not going to be easy for everyone. I don’t believe, though, that we’re going to run out of jobs. I heard an economist say, “We’re not going to run out of jobs because we’re not going to run out of problems.”

即将出现的人工智能系统将增强人类的智能，就像机械机器增强了人类的体力一样。它们不会取代人类。这并不意味着，仅仅因为分析 MRI 图像的人工智能系统能够更好地检测肿瘤，放射科医生就会失业。这将是一份截然不同的工作，一份更有趣的工作。他们将花时间做更有趣的事情，比如与患者交谈，而不是每天盯着屏幕 8 小时。

The upcoming AI systems are going to be an amplification of human intelligence in the way that mechanical machines have been an amplification of physical strength. They’re not going to be a replacement. It’s not like just because AI systems that analyze MRI images would be better at detecting tumors, then radiologists are out of a job. It’s going to be a very different job, and it’s going to be a much more interesting job. They’re going to spend their time doing more interesting things like talking to patients instead of staring at screens for 8 hours a day.

马丁·福特：但并不是每个人都是医生。很多人是出租车司机、卡车司机或快餐店员工，他们转型可能会更加困难。

MARTIN FORD: Not everyone’s a doctor, though. A lot of people are taxi drivers or truck drivers or fast food workers and they may have a harder time transitioning.

YANN LECUN：未来，产品和服务的价值将会发生改变。所有由机器完成的工作都将变得便宜很多，而所有由人类完成的工作都将变得更加昂贵。我们将为真实的人类体验付出更多，而机器可以完成的工作将变得便宜。

YANN LECUN: What’s going to happen is the value of things and services is going to change. Everything that’s by done by machine is going to get a lot cheaper, and anything that’s done by humans is going to get more expensive. We’re going to pay more for authentic human experience, and the stuff that can be done by machine is going to get cheap.

举个例子，你可以花 46 美元购买一台蓝光播放器。如果你想想蓝光播放器中采用了多少令人难以置信的先进技术，那么它的价格竟高达 46 美元，这简直是疯了。它采用了 20 年前还不存在的蓝色激光技术。它拥有一种极其精确的伺服机制，可以将激光驱动到微米级的精度。它还采用了 H.264 视频压缩和超快处理器。它采用了如此多的技术，而它之所以售价 46 美元，是因为它基本上是由机器批量生产的。现在，上网搜索手工制作的陶瓷沙拉碗，你得到的前几个搜索结果都会推荐手工制作的陶瓷碗，这是一种拥有 10,000 年历史的技术，售价约为 500 美元。为什么是 500 美元？因为它是手工制作的，你为的是体验和人与人之间的联系。你可以花一美元下载一段音乐，但如果你想去现场听一段音乐，就得花 200 美元。这是人类的体验。

As an example, you can buy a Blu-ray player for $46. If you think about how much incredibly sophisticated technology goes into a Blu-ray player, it’s insane that it costs $46. It’s got technology in the form of blue lasers that didn’t exist 20 years ago. It’s got an incredibly precise servo mechanism to drive the laser to microns of precision. It’s also got, H.264 video compression and superfast processors. It has a ridiculous amount of technology that goes in there, and it’s $46 because it’s essentially mass-produced by machines. Now, go on the web and search for a handmade ceramic salad bowl, and the first couple of hits you’re going to get are going to propose handmade ceramic bowl, a 10,000-year-old technology, for something in the region of $500. Why $500? Because it’s handmade and you’re paying for the human experience and the human connection. You can download a piece of music for a buck, but then if you want to go to a show where that music is being played live, it’s going to be $200. That’s for human experience.

事物的价值将发生变化，人们将更加重视人类体验，而自动化事物的价值将降低。出租车的价格将降低，因为它可以由人工智能系统驾驶，但由真人为您服务或真人烹饪的餐厅的价格将更高。

The value of things is going to change, with more value placed on human experience and less to things that are automated. A taxi ride is going to be cheap because it can be driven by the AI system, but a restaurant where an actual person serves you or an actual human cook creates something, is going to be more expensive.

马丁·福特：这确实假设每个人都拥有一项具有市场价值的技能或才能，我不确定这是否正确。您如何看待全民基本收入这一想法作为适应这些变化的一种方式？

MARTIN FORD: That does presume that everyone’s got a skill or talent that’s marketable, which I’m not sure is true. What do you think of the idea of a universal basic income as a way to adapt to these changes?

YANN LECUN：我不是经济学家，因此对此没有专业意见，但我采访过的每一位经济学家似乎都反对全民基本收入。他们都同意这样一个事实：由于技术进步导致不平等加剧，政府必须采取一些措施来弥补。他们都认为这与税收形式的财政政策以及财富和收入再分配有关。

YANN LECUN: I’m not an economist, so I don’t have an informed opinion on this, but every economist I talked to seemed against the idea of a universal basic income. They all agree with the fact that as a result of increased inequality brought about by technological progress, some measures have to be taken by governments to compensate. All of them believe this has to do with fiscal policy in the form of taxing, and wealth and income redistribution.

这种收入不平等在美国尤为明显，但在西欧也存在，但规模较小。法国或斯堪的纳维亚半岛的基尼系数（衡量收入不平等的指标）约为 25 或 30。在美国，基尼系数为 45，与第三世界国家处于同一水平。在美国，麻省理工学院的经济学家埃里克·布林约尔松与他的麻省理工学院同事安德鲁·麦卡菲合著了几本书，研究技术对经济的影响。他们说，自 20 世纪 80 年代里根经济学和高收入者减税以来，美国家庭的平均收入一直持平，而生产率却或多或少地持续上升。西欧没有发生过这些情况。所以，这完全是财政政策的问题。这可能是由技术进步推动的，但政府可以采取一些简单措施来弥补这种破坏，而美国政府却没有这样做。

This income inequality is something that is particularly apparent in the US, but also to a smaller scale in Western Europe. The Gini index—a measure of income inequality—of France or Scandinavia is around 25 or 30. In the US, it’s 45, and that’s the same level as third-world countries. In the US, Erik Brynjolfsson, an economist at MIT, wrote a couple of books with his colleague from MIT, Andrew McAfee, studying the impact of technology on the economy. They say that the median income of a household in America has been flat since the 1980s where we had Reaganomics and the lowering of taxes for higher incomes, whereas productivity has gone up more or less continuously. None of that occurred in Western Europe. So, it’s purely down to fiscal policy. It’s maybe fueled by technological progress, but there are easy things that governments can do to compensate for the disruption, and they’re just not doing it in the US.

马丁·福特：除了对就业市场和经济产生影响之外，人工智能还会带来哪些其他风险？

MARTIN FORD: What other risks are there, beyond the impact on the job market and economy, that come coupled with AI?

YANN LECUN：首先我想说一件我们不必担心的事情，那就是《终结者》的场景。这个想法是，我们将会以某种方式找到人工智能的秘密，创造出一种人类级别的智能，这种智能将摆脱我们的控制，然后机器人就会突然想要统治世界。统治世界的欲望与智力无关，而是与睾丸激素有关。

YANN LECUN: Let me start with one thing we should not worry about, the Terminator scenario. This idea that somehow we’ll come up with the secret to artificial general intelligence, and that we’ll create a human-level intelligence that will escape our control and all of a sudden robots will want to take over the world. The desire to take over the world is not correlated with intelligence, it’s correlated with testosterone.

今天美国政治中有很多例子，清楚地表明权力欲望与智力无关。

We have a lot of examples today in American politics, clearly illustrating that the desire for power is not correlated with intelligence.

马丁·福特：不过，尼克·博斯特罗姆提出了一个相当合理的论点。问题不在于人工智能天生就需要统治世界，而在于人工智能可能会被赋予一个目标，然后它可能会决定以一种对我们有害的方式来追求这个目标。

MARTIN FORD: There is a pretty reasoned argument, though, that Nick Bostrom, in particular, has raised. The problem is not an innate need to take over the world, but rather that an AI could be given a goal and then it might decide to pursue that goal in a way that turns out to be harmful to us.

YANN LECUN：那么，如果我们足够聪明，能够制造出通用人工智能机器，那么我们要做的第一件事就是告诉它们制造尽可能多的回形针，然后它们会把整个宇宙变成回形针？这听起来不现实。

YANN LECUN: So, somehow we’re smart enough to build artificial general intelligence machines, then the first thing we do is tell them to build as many paper clips as they can and they turn the entire universe into paper clips? That sounds unrealistic to me.

马丁·福特：我认为尼克想举一个卡通的例子。这些场景似乎都很牵强，但如果你真的在谈论超级智能，那么你就会有一台机器，它的行为方式可能让我们无法理解。

MARTIN FORD: I think Nick intends that as kind of a cartoonish example. Those kinds of scenarios all seem far-fetched, but if you are truly talking about superintelligence, then you would have a machine that might act in ways that would be incomprehensible to us.

YANN LECUN：嗯，这是目标函数设计的问题。所有这些场景都假设你会以某种方式提前设计这些机器的目标函数（内在动机），如果你做错了，它们就会做出疯狂的事情。但人类并不是这样构造的。我们的内在目标函数不是硬连线的。从某种意义上说，我们有进食、呼吸和繁殖的本能，但我们的很多行为和价值体系都是后天习得的。

YANN LECUN: Well, there is the issue of objective function design. All of those scenarios assume that somehow, you’re going to design the objective function—the intrinsic motivations—of those machines in advance, and that if you get it wrong, they’re going to do crazy things. That’s not the way humans are built. Our intrinsic objective functions are not hardwired. A piece of it is hardwired in a sense that we have the instinct to eat, breathe, and reproduce, but a lot of our behavior and value system is learned.

我们完全可以对机器做同样的事情，训练它们的价值体系，训练它们在社会中的行为举止，造福人类。这不仅仅是设计这些功能的问题，也是训练它们的问题，训练一个实体的行为举止要容易得多。我们对我们的孩子这样做，教育他们什么是对的，什么是错的，如果我们知道如何对孩子这样做，为什么我们不能对机器人或人工智能系统这样做呢？

We can very much do the same with machines, where their value system is going to be trained and we’re going to train them to essentially behave in society and be beneficial to humanity. It’s not just a problem of designing those functions but also training them, and it’s much easier to train an entity to behave. We do it with our kids to educate them in what’s right and wrong, and if we know how to do it with kids why wouldn’t we be able to do this with robots or AI systems?

显然，这里面存在一些问题，但这有点像我们还没有发明内燃机，却已经开始担心无法发明刹车和安全带。发明内燃机的问题比发明刹车和安全带要复杂得多。

Clearly, there are issues there, but it’s a bit like we haven’t invented the internal combustion engine yet and we are already worrying that we’re not going to be able to invent the brake and the safety belt. The problem of inventing the internal combustion engine is considerably more complicated than inventing brakes and safety belts.

马丁·福特：您如何看待快速起飞的情景？在这种情景下，我们会以惊人的速度不断取得进步，在不知不觉中，我们就会得到一些让我们看起来就像老鼠或昆虫的东西。

MARTIN FORD: What do you think of the fast takeoff scenario, where you have recursive improvement that happens at an extraordinary rate, and before you know it, we’ve got something that makes us look like a mouse or an insect in comparison?

YANN LECUN：我绝对不相信这一点。显然，人工智能会不断进步，而且毫无疑问，机器越智能，它们就越能帮助我们设计下一代产品。事实已经如此，而且这种趋势还会加速。

YANN LECUN: I absolutely do not believe in that. Clearly there’s going to be continuous improvement, and certainly, the more intelligent machines become, the more they’re going to help us design the next generation. It’s already the case, and it’s going to accelerate.

有某种微分方程控制着技术进步、经济、资源消耗、通信、技术复杂化等所有东西。这个方程中有一大堆摩擦项，而奇点或快速起飞的支持者却完全忽略了这些摩擦项。每个物理过程在某个时刻都必须达到饱和状态，至少会耗尽资源。所以，我不相信快速起飞。有人会弄清楚 AGI 的秘密，然后突然之间，我们的机器就会从像老鼠一样聪明变成像猩猩一样聪明，然后一周后它们就比我们聪明了，一个月后，它们就比我们聪明多了，这是一种谬论。

There is some sort of differential equation that governs the progress of technology, the economy, consumption of resources, communication, the sophistication of technology, and all that stuff. There’s a whole bunch of friction terms in this equation that is completely ignored by the proponent of singularity or fast takeoff. Every physical process at some point has to saturate, by exhausting resources if nothing else. So, I don’t believe in a fast takeoff. It’s a fallacy that someone will figure out the secret to AGI, then all of a sudden, we’re going to go from machines that are as intelligent as a rat to some that are as intelligent as an orangutan, and then a week later they are more intelligent than us, and then a month later, way more intelligent.

也没有理由相信，如果机器比单个人类聪明很多，它就能够完全超越单个人类。人类可能会被极其愚蠢的病毒杀死，但它们专门用来杀死我们。

There’s also no reason necessarily to believe that being way more intelligent than a single human will allow a machine to be completely superior to a single human. Humans can get killed by viruses that are extremely stupid, but they are specialized to kill us.

如果我们能够构建一个具有这种通用智能的人工智能系统，那么我们可能也能构建一个更专业的智能，旨在摧毁第一个智能。它会更有效地杀死 AGI，因为更专业的机器比通用机器更有效率。我只是认为每个问题都有自己的内置解决方案。

If we can build an artificial intelligence system that has general intelligence in that sense, then we can probably also build a more specialized intelligence designed to destroy the first one. It would be much more efficient at killing the AGI because more specialized machines are more efficient than general ones. I just think that every issue has its own solution built in.

马丁·福特：那么，未来十年或二十年我们真正应该担心的是什么呢？

MARTIN FORD: So, what should we legitimately be worried about in the next decade or two?

YANN LECUN：经济混乱显然是一个问题。这不是一个没有解决方案的问题，但这是一个存在巨大政治障碍的问题，特别是在美国这样的文化中，收入和财富再分配在文化上并不被接受。还有一个问题是如何传播这项技术，以便它不仅能造福发达国家，还能让全世界共享。

YANN LECUN: Economic disruption is clearly an issue. It’s not an issue without a solution, but it’s an issue with considerable political obstacles, particularly in cultures like the US where income and wealth redistribution are not something that’s culturally accepted. There is an issue of disseminating the technology so that it doesn’t only profit the developed world, but it’s shared across the world.

权力集中。目前，人工智能研究非常公开，但目前只有相对少数的公司在广泛部署。它还需要一段时间才能被更广泛的经济领域所使用，这就是权力的重新分配。这将在某些方面影响世界，可能是积极的，也可能是消极的，我们需要确保它是积极的。

There is a concentration of power. Currently, AI research is very public and open, but it’s widely deployed by a relatively small number of companies at the moment. It’s going to take a while before it’s used by a wider swath of the economy and that’s a redistribution of the cards of power. That will affect the world in some ways, it may be positive but it may also be negative, and we need to ensure that it’s positive.

我认为技术进步的加速和人工智能的出现将促使政府加大对教育的投资，尤其是继续教育，因为人们将不得不学习新的工作。这是需要处理的颠覆的一个真正方面。这不是没有解决方案的问题，只是人们必须意识到它的存在，才能解决它。

I think the acceleration of technological progress and the emergence of AI is going to prompt governments to invest more massively into education, particularly continuous education because people are going to have to learn new jobs. That’s a real aspect of the disruption that needs to be dealt with. It’s not something that doesn’t have a solution, it’s just a problem that people have to realize exists in order for them to solve it.

如果政府连全球变暖等既定的科学事实都不相信，他们怎么会相信这种东西呢？这类问题有很多，包括偏见和公平问题。如果我们使用监督学习来训练我们的系统，它们就会反映出数据中的偏见，那么你如何确保它们不会延续偏见的现状呢？

If you have a government that doesn’t even believe in established scientific facts like global warming, how can they believe in this kind of stuff? There are a lot of issues of this type, including ones in the area of bias and equity. If we use supervised learning to train our systems, they’re going to reflect the biases that are in the data, so how can you make sure they don’t prolong the status quo in terms of biases?

马丁·福特：问题在于，偏见被封装在数据中，因此机器学习算法自然会获得它们。人们希望修复算法中的偏见比修复人类中的偏见要容易得多。

MARTIN FORD: The problem there is that the biases are encapsulated in the data so that a machine learning algorithm would naturally acquire them. One would hope that it might be much easier to fix bias in an algorithm than in a human.

YANN LECUN：当然。我其实对这个方面非常乐观，因为我认为减少机器的偏见确实比减少人类的偏见要容易得多。人类的偏见是很难纠正的。

YANN LECUN: Absolutely. I’m actually quite optimistic in that dimension because I think it would indeed be a lot easier to reduce bias in a machine than it currently is with people. People are biased in ways that are extremely difficult to fix.

马丁·福特：您是否担心军事应用，比如自主武器？

MARTIN FORD: Do you worry about military applications, like autonomous weapons?

YANN LECUN：是也不是。是的，因为人工智能技术当然可以用来制造武器，但有些人，比如 Stuart Russell，将潜在的新一代人工智能武器描述为大规模杀伤性武器，我完全不同意这种观点。

YANN LECUN: Yes and no. Yes, because of course AI technology can be used for building weapons, but some people, like Stuart Russell, have characterized a potential new generation of AI-powered weapons as weapons of mass destruction and I completely disagree with that.

我认为军队使用人工智能技术的方式恰恰相反。军队称之为外科手术式行动。你不会投下一颗炸弹摧毁整栋建筑，而是派出无人机让你想要捕获的人进入睡眠状态；这可能是非致命的。

I think the way that militaries are going to use AI technology is exactly the opposite. It’s for what the military calls, surgical actions. You don’t drop a bomb that destroys an entire building, you send in your drone that just puts the person you are interested in capturing to sleep; it could be non-lethal.

当核武器发展到那个地步时，军队看起来就更像警察了。从长远来看，这样做好吗？我想没人能猜到。它的破坏力比核武器小——不可能比核武器更大！

When it gets to that point, it makes the military look more like police. Is that good in the long term? I don’t think anyone can guess. It’s less destructive than nukes—it can’t be more destructive than nukes!

马丁·福特：您是否担心在人工智能方面与中国展开竞争？中国有十多亿人口，因此拥有更多数据，同时隐私限制也更少。这会让他们在未来的发展中占据优势吗？

MARTIN FORD: Do you worry about a race with China in terms of advancing artificial intelligence? They have over a billion people, so they have got more data and along with that, fewer constraints on privacy. Is that going to give them an advantage in moving forward?

YANN LECUN：我不这么认为。我认为目前科学的进步并不依赖于数据的广泛可用性。中国可能有10多亿人口，但真正参与技术和研究的人的比例实际上相对较小。

YANN LECUN: I don’t think so. I think currently progress in the science is not conditioned on the wide availability of data. There may be more than 1 billion people in China, but the proportion of people who are actually involved in technology and research is actually relatively small.

毫无疑问，它会发展，中国确实在朝这个方向前进。我认为，一段时间后，政府的作风和教育类型可能会扼杀创造力。不过，中国也有一些出色的工作，那里有一些非常聪明的人，他们将为这个领域做出贡献。

There’s no question that it will grow, China is really progressing in that direction. I think the style of government and the type of education they have may be stifling for creativity after a while. There is good work coming out of China, though, with some very smart people there, and they’re going to make contributions to this field.

20 世纪 80 年代，西方曾有过同样的担忧，担心日本技术会占领西方市场，这种情况持续了一段时间，然后逐渐饱和。后来是韩国，现在是中国。未来几十年，中国社会将发生重大变化，这可能会彻底改变现状。

There was the same kind of fear of the West being overrun by Japanese technology in the 1980s, and it happened for a while and then it kind of saturated. Then it was the Koreans, and now it’s the Chinese. There are going to be big mutations in Chinese society that will have to happen over the next few decades that will probably change the situation completely.

马丁·福特：您认为人工智能需要某种程度的监管吗？政府是否应该对您所从事的研究和所构建的系统进行监管？

MARTIN FORD: Do you think that AI needs to be regulated at some level? Is there a place for government regulation for the kind of research you’re doing and the systems that you’re building?

YANN LECUN：虽然我认为目前监管人工智能研究没有任何意义，但我确实认为监管应用是必要的。不是因为它们使用人工智能，而是因为它们的应用领域。

YANN LECUN: While I don’t think there is any point in regulating AI research at the moment, I do think there is certainly a need for regulating applications. Not because they use AI, but because of the domain of applications that they are.

以药物设计中人工智能的使用为例，你总是想规范药物的测试方式、部署方式和使用方式。事实已经如此。以自动驾驶汽车为例：汽车受到监管，道路安全法规也十分严格。当然，这些应用领域可能需要调整现有法规，因为人工智能将占据主导地位。

Take the use of AI in the context of drug design; you always want to regulate how drugs are being tested, how they are deployed, and how they are used. It’s already the case. Take self-driving cars: cars are regulated, and there are strict road safety regulations. Certainly, those are application areas where existing regulations might need to be tweaked because AI is going to become preponderant.

然而，我认为目前没有必要对人工智能进行监管。

However, I don’t see any need for the regulation of AI at the moment.

马丁·福特：那么，我认为你非常不同意伊隆·马斯克所使用的言论？

MARTIN FORD: So, I assume you disagree quite strongly with the kind of rhetoric Elon Musk has been using?

YANN LECUN：哦，我完全不同意他的观点。我和他谈过几次，但我不知道他的观点从何而来。他是个非常聪明的人，我对他的一些项目感到敬畏，但我不确定他的动机是什么。他想拯救人类，所以也许他需要另一个生存威胁。我认为他真的很担心，但我们中没有人能够说服他，博斯特罗姆式的硬起飞情景不会发生。

YANN LECUN: Oh, I completely and absolutely disagree with him. I’ve talked to him several times, but I don’t know where his views are coming from. He’s a very smart guy and I’m in awe of some of his projects, but I’m not sure what his motivation is. He wants to save humanity, so maybe he needs another existential threat for it. I think he is genuinely worried, but none of us have been able to convince him that Bostrom-style, hard take-off scenarios are not going to happen.

马丁·福特：你总体上是个乐观主义者吗？你相信人工智能的好处会超过坏处吗？

MARTIN FORD: Are you an optimist overall? Do you believe that the benefits of AI are going to outweigh the downsides?

YANN LECUN：是的，我同意这一点。

YANN LECUN: Yes, I would agree with that.

马丁·福特：您认为它会在哪些方面带来最大的益处？

MARTIN FORD: In what areas do you think it will bring the most benefits?

YANN LECUN：我真的希望我们能找到让机器像人类婴儿和动物一样学习的方法。这是我未来几年的科学计划。我还希望我们能在资助所有这些研究的人感到厌倦之前取得一些令人信服的突破，因为这就是过去几十年发生的事情。

YANN LECUN: Well, I really hope that we figure out the way to get machines to learn like baby humans and animals. That’s my scientific program for the next few years. I also hope we’re going to make some convincing breakthrough before the people funding all this research get tired, because that’s what happened in previous decades.

马丁·福特：你曾警告说，人工智能被过度炒作，这甚至可能导致又一个“人工智能寒冬”。你真的认为存在这种风险吗？深度学习已成为谷歌、Facebook、亚马逊、腾讯和所有其他非常富有的公司商业模式的核心。因此，很难想象对这项技术的投资会大幅下降。

MARTIN FORD: You’ve warned that AI is being overhyped and that this might even lead to another “AI Winter.” Do you really think there’s a risk of that? Deep learning has become so central to the business models of Google, Facebook, Amazon, Tencent, and all these other incredibly wealthy corporations. So, it seems hard to imagine that investment in the technology would fall off dramatically.

YANN LECUN：我不认为我们会看到像以前那样的人工智能寒冬，因为人工智能周围有一个很大的产业，而且有真正的应用正在为这些公司带来真正的收入。

YANN LECUN: I don’t think we’re going to see an AI winter in the way we saw before because there is a big industry around it and there are real applications that are bringing real revenue to these companies.

仍有大量投资，例如，人们希望自动驾驶汽车在未来五年内投入使用，医学成像将发生彻底的革命。这些可能是未来几年最明显的影响，包括医药和医疗保健、交通和信息访问。

There’s still a huge amount of investment, with the hope that, for example, self-driving cars are going to be working in the next five years and that medical imaging is going to be radically revolutionized. Those are probably going to be the most visible effects over the next few years, medicine and health care, transportation, and information access.

虚拟助手是另一个例子。它们目前用处不大，因为它们是手写脚本。它们没有任何常识，也无法真正理解你告诉它们的内容。问题是，在我们获得不会令人沮丧的虚拟助手之前，我们是否需要解决 AGI 问题，或者我们是否可以在那之前取得更多持续进展。现在，我不知道。

Virtual assistants are another case. They are only mildly useful today because they’re kind of scripted by hand. They don’t have any common sense, and they don’t really understand what you tell them at a deep level. The question is whether we need to solve the AGI problem before we get virtual assistants that are not frustrating, or whether we can make more continuous progress before that. Right now, I don’t know.

但一旦这项技术问世，人们之间的互动方式以及人们与数字世界的互动方式将会发生很大变化。如果每个人都拥有一个拥有人类水平智能的个人助理，那将会带来巨大的改变。

When that becomes available, though, that’s going to change a lot of how people interact with each other and how people interact with the digital world. If everyone has a personal assistant that has human-level intelligence, that’s going to make a huge difference.

不知道你看过《她》这部电影吗？这部电影对可能发生的事情的描述从某些方面来说还不错。在所有关于人工智能的科幻电影中，它可能是最不荒谬的一部。

I don’t know if you’ve seen the movie Her? It’s not a bad depiction in some ways of what might happen. Among all the sci-fi movies on AI, it’s probably one of the least ridiculous.

我认为，随着硬件的进步，许多与人工智能相关的技术将广泛普及。现在人们正在努力开发低功耗廉价硬件，这些硬件可以装入智能手机或吸尘器中，以 100 毫瓦的功率运行卷积网络，芯片只需 3 美元即可购买。这将极大地改变我们周围的世界。

I think a lot of AI-related technology is going to be widely available in the hands of people because of hardware progress. There’s a lot of effort now to develop low-power and cheap hardware that can fit in your smartphone or your vacuum cleaner that can run a convolutional network on 100 milliwatts of power, and the chip can be bought for 3 bucks. That’s going to change a lot of how the world around us works.

您的吸尘器不再会在房间里随意走动，而是能够看到它需要去的地方，您的割草机也能够在不碾过花坛的情况下修剪草坪。不仅仅是您的汽车可以自动驾驶。

Instead of going randomly around your room, your vacuum cleaner is now going to be able to see where it needs to go, and your lawnmower is going to be able to mow your lawn without running over your flowerbeds. It’s not just your car that will drive itself.

它还可能对环境产生有趣的影响，比如野生动物监测。人工智能将进入每个人的手中，因为专门用于深度学习的硬件技术正在取得进步，而这一技术将在未来 2 到 3 年内出现。

It might also have interesting environmental consequences, like wildlife monitoring. AI is going to be in the hands of everyone because of progress in hardware technology that is specialized for deep learning, and that’s coming in the next 2 or 3 years.

YANN LECUN 是 Facebook 的副总裁兼首席人工智能科学家，也是纽约大学的计算机科学教授。Yann 与 Geoff Hinton 和 Yoshua Bengio 一起，是所谓的“加拿大黑手党”的成员——这三位研究人员的努力和坚持直接导致了深度学习神经网络的当前革命。

YANN LECUN is a Vice President and Chief AI Scientist at Facebook, as well as a professor of computer science at New York University. Along with Geoff Hinton and Yoshua Bengio, Yann is part of the so-called “Canadian Mafia”—the trio of researchers whose effort and persistence led directly to the current revolution in deep learning neural networks.

在加入 Facebook 之前，Yann 曾在 AT&T 的贝尔实验室工作，在那里他因开发卷积神经网络而备受赞誉，这是一种受大脑视觉皮层启发的机器学习架构。Yann 使用卷积神经网络开发了一种手写识别系统，该系统被广泛用于 ATM 和银行读取支票上的信息。近年来，由速度越来越快的计算机硬件驱动的深度卷积网络彻底改变了计算机图像识别和分析。

Prior to joining Facebook, Yann worked at AT&T’s Bell Labs, where he is credited with developing convolutional neural networks—a machine learning architecture inspired by the brain’s visual cortex. Yann used convolutional neural nets to develop a handwriting recognition system that became widely used in ATMs and at banks to read the information on checks. In recent years, deep convolutional nets, powered by ever faster computer hardware, have revolutionized computer image recognition and analysis.

Yann 于 1987 年获得巴黎高等电子技术与电子工程师学院 (ESIEE) 的电气工程师文凭，​​并于 1987 年获得皮埃尔和玛丽居里大学的计算机科学博士学位。后来，他在多伦多大学的 Geoff Hinton 实验室担任博士后研究员。他于 2013 年加入 Facebook，建立并运营总部位于纽约市的 Facebook AI 研究 (FAIR) 组织。

Yann received an Electrical Engineer Diploma from Ecole Superieure d’Ingenieurs en Electrotechnique et Electronique (ESIEE) in Paris, and a PhD in Computer Science from Universite Pierre et Marie Curie in 1987. He later worked as a post-doctoral researcher in Geoff Hinton’s lab at the University of Toronto. He joined Facebook in 2013 to establish and run the Facebook AI Research (FAIR) organization, headquartered in New York City.

斯坦福大学计算机科学教授 谷歌云首席科学家

PROFESSOR OF COMPUTER SCIENCE, STANFORD CHIEF SCIENTIST, GOOGLE CLOUD

李飞飞是斯坦福大学计算机科学教授，也是斯坦福人工智能实验室 (SAIL) 主任。飞飞从事计算机视觉和认知神经科学领域的工作，她构建了智能算法，使计算机和机器人能够观察和思考，这些算法的灵感来自人类大脑在现实世界中的工作方式。飞飞是 Google Cloud 的 AI 和机器学习首席科学家，她致力于推动 AI 的发展和普及。飞飞是人工智能多元化和包容性的坚定支持者，并共同创立了 AI4ALL，这是一个旨在吸引更多女性和来自代表性不足群体的人进入该领域的组织。

Fei-Fei Li is Professor of Computer Science at Stanford University, and Director of the Stanford Artificial Intelligence Lab (SAIL). Working in areas of computer vision and cognitive neuroscience, Fei-Fei builds smart algorithms that enable computers and robots to see and think, inspired by the way the human brain works in the real world. Fei-Fei is Chief Scientist, AI and Machine Learning at Google Cloud, where she works to advance and democratize AI. Fei-Fei is a strong proponent of diversity and inclusion in artificial intelligence and co-founded AI4ALL, an organization to attract more women and people from underrepresented groups into the field.

马丁·福特：我们来谈谈你的职业发展轨迹吧。你是怎么开始对人工智能感兴趣的？这又是如何让你获得斯坦福大学的现任职位的？

MARTIN FORD: Let’s talk about your career trajectory. How did you first become interested in AI, and how did that lead to your current position at Stanford?

李飞飞：我一直都是 STEM 的学生，所以科学一直吸引着我，我尤其喜欢物理。我考入普林斯顿大学主修物理学，学习物理的一个副产品是我对宇宙的基本原理着迷。诸如宇宙从何而来？存在的意义是什么？宇宙将走向何方？这些都是人类好奇心的基本探索。

FEI-FEI LI: I’ve always been something of a STEM student, so the sciences have always appealed to me, and in particular I love physics. I went to Princeton University where I majored in Physics, and a by-product of studying physics is that I became fascinated by the fundamentals of the universe. Questions like, where does the universe come from? What does it mean to exist? Where is the universe going? The fundamental quest of human curiosity.

我在研究中注意到一件非常有趣的事情：自 20 世纪初以来，我们见证了现代物理学的伟大觉醒，这要归功于爱因斯坦和勋伯格等人，他们在生命的尽头不仅对宇宙的物质感兴趣，而且对生命、生物学以及存在的基本问题也非常感兴趣。我也对这些问题非常着迷。当我开始研究时，我意识到我对生活的真正兴趣不是发现物质，而是理解智慧——它定义了人类的生命。

In my research I noticed something really interesting: since the beginning of the 20th century, we’ve seen a great awakening of modern physics, due to the likes of Einstein and Schoenberg, who towards the end of their lives became fascinated not only by the physical matter of the universe but by life, and biology, and by the fundamental questions of being. I became very fascinated by these questions as well. When I started to study, I realized that my real interest in life is not to discover physical matters but to understand intelligence—which defines human life.

马丁·福特：这是您在中国的时候吗？

MARTIN FORD: Was this when you were in China?

李飞飞：我在美国普林斯顿大学物理系学习时，开始对人工智能和神经科学产生兴趣。当我在那里申请博士学位时，我非常幸运，直到今天，我所做的事情——既研究神经科学，又研究人工智能——仍然是一种罕见的组合。

FEI-FEI LI: I was in the US, at Princeton Physics, when my intellectual interest in AI and neuroscience began. When I applied for a PhD there I was very lucky, and to this day, it’s still a bit of a rare combination to do what I did—which was both neuroscience and AI.

马丁·福特：那么您认为研究这两个领域而不是仅仅专注于计算机科学驱动的方法是一个重要的优势吗？

MARTIN FORD: Do you think then that it’s an important advantage to study both of those fields rather than to focus exclusively on a computer-science-driven approach?

李飞飞：我认为这给了我一个独特的视角，因为我认为自己是一名科学家，所以当我接触人工智能时，驱动我的是科学假设和科学探索。人工智能领域是关于思考机器、让机器变得智能，我喜欢研究征服机器智能的核心问题。

FEI-FEI LI: I think it gives me a unique angle because I consider myself a scientist, and so when I approach AI, what drives me is scientific hypotheses and the scientific quest. The field of AI is about thinking machines, making machines intelligent, and I like to work on problems at the core of conquering machine intelligence.

由于我有认知神经科学背景，所以我从算法的角度和详细的建模角度来看待问题。因此，我发现大脑和机器学习之间的联系非常有趣。我还经常思考推动人工智能进步的受人类启发的任务：我们的自然智能必须通过进化来解决现实世界的任务。我的背景让我有了独特的视角和方法来处理人工智能。

Coming from a cognitive neuroscience background, I take an algorithmic point of view, and a detailed modeling point of view. So, I find the connection between the brain and machine learning fascinating. I also think a lot about human-inspired tasks that drive AI advances: the real-world tasks that our natural intelligence had to solve through evolution. My background has in this way given me a unique angle and approach to working with AI.

马丁·福特：你的重点确实是计算机视觉，你指出，从进化的角度来看，眼睛的发育很可能导致大脑本身的发育。大脑提供解释图像的计算能力，因此理解视觉可能是通往智能的大门。我这样想对吗？

MARTIN FORD: Your focus has really been on computer vision, and you’ve made the point that, in evolutionary terms, the development of the eye likely led to the development of the brain itself. The brain was providing the compute power to interpret images, and so maybe understanding vision is the gateway to intelligence. Am I correct in that line of thinking?

李飞飞：是的，你说得对。语言当然是人类智能的重要组成部分：与言语、触觉、决策和推理一样。但视觉智能却嵌入在所有这些事物之中。

FEI-FEI LI: Yes, you’re right. Language is a huge part of human intelligence, of course: along with speech, tactile awareness, decision-making, and reasoning. But visual intelligence is embedded in all of these things.

如果你看看大自然是如何设计我们大脑的，就会发现人类大脑的一半与人类智能有关，而人类智能与运动系统、决策、情感、意图和语言密切相关。人类大脑并非只是偶然识别孤立的物体；这些功能是人类智能不可或缺的一部分。

If you look at the way nature designed our brain, half of the human brain is involved in human intelligence, and that human intelligence is intimately related to a motor system, to decision-making, to emotion, to intention, and to language. The human brain does not just happen to recognize isolated objects; these functions are an integral part of what deeply defines human intelligence.

马丁·福特：您能概括一下您在计算机或机器视觉领域所做的一些工作吗？

MARTIN FORD: Could you sketch out some of the work you’ve done in computer or machine vision?

李飞飞：在 21 世纪的第一个十年，物体识别是计算机视觉领域一直努力追求的圣杯。物体识别是所有视觉的基础。作为人类，如果我们睁开眼睛环顾周围环境，我们几乎可以识别我们看到的每个物体。识别对于我们能够探索世界、理解世界、交流世界和做世界上的事情至关重要。物体识别是计算机视觉领域一个非常崇高的圣杯，当时我们正在使用机器学习等工具。

FEI-FEI LI: During the first decade of the 21st century, object recognition was the holy grail that the field of computer vision was working on. Object recognition is a building block for all vision. As humans, if we open our eyes and look around our environment, we recognize almost every object we look at. Recognition is critically important for us to be able to navigate the world, understand the world, communicate about the world, and do things in the world. Object recognition was a very lofty holy grail in computer vision, and we were using tools such as machine learning at that time.

然后在 2000 年代中期，当我从博士生转为教授时，很明显计算机视觉领域陷入了困境，机器学习模型并没有取得巨大进步。当时，整个国际社会都在用大约 20 种不同的物体对自动识别任务进行基准测试。

Then in the mid-2000s, as I transitioned from a PhD student to become a professor, it became obvious that computer vision as a field was stuck, and that the machine learning models were not making huge progress. Back then, the whole international community was benchmarking autorecognition tasks with around 20 different objects.

因此，我和我的学生和同事开始深入思考如何实现质的飞跃。我们开始意识到，仅仅解决涉及 20 个物体的如此小规模的问题并不足以实现物体识别的崇高目标。我深受人类认知的启发，也深受任何儿童的发展故事的启发，儿童在最初几年的发展过程中需要大量的数据。儿童对他们的世界进行大量的实验，观察世界，并吸收它。巧合的是，就在这个时候，互联网蓬勃发展成为一种全球现象，提供了大量大数据。

So, along with my students and collaborators, we started thinking deeply about how we might make a quantum leap forward. We began to see that it was just not going to be sufficient for us to work with such a small-scale problem involving 20 objects to reach the lofty goal of object recognition. I was very much inspired by human cognition at this point, and the developmental story of any child, where the first few years of development involves a huge amount of data. Children engage in a huge amount of experimenting with their world, seeing the world, and just taking it in. Coincidentally, at was just at this time that the internet had boomed into a global phenomenon that provided a lot of big data.

我想做一个相当疯狂的项目，把我们在互联网上能找到的所有图片整理成对人类有意义的概念，并给这些图片贴上标签。结果，这个疯狂的想法变成了 ImageNet 项目，将 1500 万张图片整理成 22,000 个标签。

I wanted to do a pretty crazy project that would take all the pictures we could find on the internet, organize them into concepts that mattered to humans, and label those images. As it turned out, this crazy idea turned into the project called ImageNet, with 15 million images organized into 22,000 labels.

我们立即向全世界开放了 ImageNet 的源代码，因为直到今天我仍然相信技术的民主化。我们向全世界发布了全部 1500 万张图片，并开始举办国际竞赛，让研究人员研究 ImageNet 问题：不是研究微小的问题，而是研究对人类和应用程序至关重要的问题。

We immediately open-sourced ImageNet for the world, because to this day I believe in the democratization of technology. We released the entire 15 million images to the world and started to run international competitions for researchers to work on the ImageNet problems: not on the tiny small-scale problems but on the problems that mattered to humans and applications.

时间快进到 2012 年，我认为我们看到了许多人物体识别领域的转折点。2012 年 ImageNet 竞赛的获胜者创造了一种融合了 ImageNet、GPU 计算能力和卷积神经网络的算法。Geoffrey Hinton 写了一篇开创性的论文，对我来说，这是实现物体识别圣杯的第一阶段。

Fast-forward to 2012, and I think we see the turning point in object recognition for a lot of people. The winner of the 2012 ImageNet competition created a convergence of ImageNet, GPU computing power, and convolutional neural networks as an algorithm. Geoffrey Hinton wrote a seminal paper that, for me, was Phase One in achieving the holy grail of object recognition.

马丁·福特：你继续这个项目了吗？

MARTIN FORD: Did you continue this project?

李飞飞：在接下来的两年里，我致力于进一步提高物体识别能力。如果我们再看看人类的发展，婴儿从牙牙学语开始，说几个词，然后开始造句。我有一个两岁的女儿和一个六岁的儿子。两岁的孩子能说很多句子，这是一个巨大的发展进步，这是人类作为智能体和动物所做的事情。受人类发展的启发，我开始研究如何让计算机在看到图片时说出句子，而不仅仅是给椅子或猫贴上标签。

FEI-FEI LI: For the next two years, I worked on taking object recognition a step further. If we again look at human development, babies start by babbling, a few words, and then they start making sentences. I have a two-year-old daughter and a six-year-old son. The two-year-old is making a lot of sentences, which is huge developmental progress, something that humans do as intelligent agents and animals. Inspired by this human development, I started working on the problem of how to enable computers to speak sentences when they see pictures, rather than just labeling a chair or a cat.

我们用深度学习模型研究这个问题已有几年了。2015 年，我在 TED2015 大会上谈到了这个项目。我演讲的题目是“我们如何教计算机理解图片”，我讨论了如何让计算机能够理解图像的内容，并将其总结为人类自然语言的句子，然后进行交流。

We were working on this problem using deep learning models for a few years. In 2015, I talked about the project at the TED2015 conference. The title of my talk was How we’re teaching computers to understand pictures, and I discussed enabling computers to be able to understand the content of an image and summarize it in a human, natural-language sentence which could then be communicated.

马丁·福特：算法的训练方式与人类婴儿或幼儿的训练方式截然不同。大多数情况下，儿童不会获得标记数据——他们只是自己摸索。即使你指着一只猫说，“看，那是一只猫”，你也不必重复说上十万遍。一两次可能就够了。人类从我们在世界上遇到的非结构化实时数据中学习的方式与现在人工智能的监督学习方式有着相当大的区别。

MARTIN FORD: The way algorithms are trained is quite different from what happens with a human baby or young child. Children for the most part are not getting labeled data—they just figure things out. And even when you point to a cat and say, “look there’s a cat,” you certainly don’t have to do that a hundred thousand times. Once or twice is probably enough. There’s a pretty remarkable difference in terms of how a human being can learn from the unstructured, real-time data we meet in the world, versus the supervised learning that’s done with AI now.

李飞飞：你说得完全正确，这就是为什么作为一名人工智能科学家，我每天醒来都兴奋不已，因为有太多工作要做。有些工作受到了人类的启发，但很大一部分工作根本不像人类。正如你所说，如今神经网络和深度学习的成功主要涉及监督模式识别，这意味着与一般的人类智能相比，它的能力非常有限。

FEI-FEI LI: You totally nailed it, and this is why as an AI scientist I wake up so excited every day because there’s so much to work with. Some part of the work has inspiration from humans, but a large part of the work does not resemble humans at all. As you say, the success today of neural networks and deep learning mostly involve supervised pattern recognition, which means that it’s a very narrow sliver of capabilities compared to general human intelligence.

今年，我在 Google 的 I/O 大会上发表了演讲，再次以我两岁的女儿为例。几个月前，我在婴儿监视器上看到她通过学习系统中的裂缝（一条逃离婴儿床的潜在路径）从婴儿床中逃脱。我看到她打开了睡袋，我特别改造了睡袋，以防止她打开睡袋自己逃出。这种与视觉运动协调的智能、计划、推理、情感、意图和坚持，在目前的人工智能中确实无处不在。我们还有很多工作要做，认识到这一点非常重要。

I gave a talk at Google’s I/O conference this year, where I was again using the example of my two-year-old daughter. A couple of months ago, I watched her on a baby monitor escape from her crib by learning the cracks in the system, a potential path to escape from the crib. I saw her open her sleeping bag, which I had particularly modified in order to prevent her from opening and get herself out. That kind of coordinated intelligence to a visual motor, planning, reasoning, emotion, intention, and persistence, is really nowhere to be seen in our current AI. We’ve got a lot of work to do, and it’s really important to recognize that.

马丁·福特：您认为未来有可能出现突破，让计算机能够像孩子一样学习吗？人们是否正在积极研究如何解决这个问题？

MARTIN FORD: Do you think there will likely be breakthroughs that allow computers to learn more like children? Are people actively working on how to solve this problem?

李飞飞：肯定有人在研究这个问题，尤其是在研究界。我们很多人都在研究下一个地平线问题。在我斯坦福的实验室里，我们正在研究机器人学习问题，其中人工智能通过模仿来学习，这比通过监督标签学习要自然得多。

FEI-FEI LI: There are absolutely people working on that, especially within the research community. A lot of us are working on the next horizon problem. In my own lab at Stanford, we are working on robotic learning problems where the AI is learning by imitation, which is much more natural than learning by supervised labels.

小时候，我们观察其他人如何做事，然后我们自己也效仿；因此，该领域现在开始涉及逆向强化学习算法和神经编程算法。有很多新的探索，DeepMind 正在做这些。谷歌大脑正在做这些；斯坦福大学正在做这些；麻省理工学院也在做这些。鉴于全球在这一领域的投资数量惊人，我非常希望在我们有生之年能看到更多的人工智能突破。我们还看到研究界在研究监督学习以外的算法方面做出了许多努力。

As kids, we watch how other humans do things and then we do it; so, the field is now starting to get into inverse reinforcement learning algorithms, and neuro-programming algorithms. There is a lot of new exploration, and DeepMind is doing that. Google Brain is doing that; Stanford is doing that; and MIT is doing that. I’m very hopeful that in our lifetime we’ll be seeing a lot more AI breakthroughs, given the incredible amount of global investment in this area. We also see a lot of effort in the research community to look at algorithms beyond supervised learning.

预测突破何时到来要困难得多。作为一名科学家，我学会了不要预测科学突破，因为它们是偶然出现的，而且是历史因素汇聚而成的。但鉴于全球在这一领域的巨额投资，我非常有希望在我们有生之年看到更多人工智能突破。

Dating when a breakthrough will come, is much harder to predict. I learned, as a scientist, not to predict scientific breakthroughs, because they come serendipitously, and they come when a lot of ingredients in history converge. But I’m very hopeful that in our lifetime we’ll be seeing a lot more AI breakthroughs given the incredible amount of global investment in this area.

马丁·福特：我知道您是 Google Cloud 的首席科学家。我在演讲时总是强调的一点是，人工智能和机器学习将像一种公用设施一样——几乎像电力一样——几乎可以在任何地方部署。在我看来，将人工智能集成到云端是让这项技术普遍可用的第一步。这符合您的愿景吗？

MARTIN FORD: I know you’re the chief scientist for Google Cloud. A point that I always make when I give presentations is that AI and machine learning are going to be like a utility—almost like electricity—something that can be deployed almost anywhere. It seems to me that integrating AI into the cloud is one of the first steps toward making the technology universally available. Is that in line with your vision?

李飞飞：作为一名教授，我们每七八年都会有休假，离开大学几年，去探索不同的工作或让自己重焕活力。两年前，我非常确定我想加入一个真正普及人工智能技术的行业，因为人工智能已经发展到一定的程度，一些目前正在发挥作用的技术，如监督学习和模式识别，正在为社会做出贡献。就像你说的，如果你想传播人工智能这样的技术，最好和最大的平台就是云，因为在人类发明的任何平台上，没有其他计算能覆盖如此多的人。单是谷歌云，在任何时候，都在赋能、帮助或服务数十亿人。

FEI-FEI LI: As a professor, every seven or eight years there is a built-in encouragement for sabbaticals where we leave the university for a couple of years to explore a different line of work or to refresh yourself. Two years ago, I was very sure that I wanted to join an industry to really democratize AI technologies, because AI has advanced to a point where some of the technology that is now working, like supervised learning and pattern recognition, is doing good things for society. And like you say, if you think about disseminating technology like AI, the best and biggest platform is a cloud because there’s no other computing on any platform which humanity has invented that reaches as many people. Google Cloud alone, at any moment, is empowering, helping, or serving billions of people.

因此，我很高兴被邀请担任 Google Cloud 的首席科学家，我们的使命是让人工智能普及大众。我们的使命是创造能够赋能企业和合作伙伴的产品，然后听取客户的反馈并与他们密切合作以改进技术本身。这样，我们就可以实现人工智能普及和人工智能进步之间的闭环。我负责云人工智能的研究部分以及云人工智能的产品，我们自 2017 年 1 月以来一直在这里工作。

I was therefore very happy to be invited as chief scientist of Google Cloud, where the mission is to democratize AI. This is about creating products that empower businesses and partners, and then taking the feedback from customers and working with them closely to improve the technology itself. This way we can close that loop between the democratization of AI and the advancement AI. I’m overseeing both the research part of cloud AI as well as the product of cloud AI, and we’ve been here since January 2017.

我们正在做的一个例子是我们创建的产品 AutoML。这是市场上独一无二的产品，它尽可能地降低了人工智能的准入门槛，这样人工智能就可以提供给那些不从事人工智能的人。客户痛点在于，许多企业需要定制模型来帮助他们解决自己的问题。所以，在计算机视觉的背景下，如果我是零售商，我可能需要一个模型来识别我的标志。如果我是《国家地理》杂志，我可能需要一个模型来识别野生动物。如果我在农业行业工作，我可能需要一个模型来识别苹果。人们有各种各样的用例，但并不是每个人都有创造人工智能的机械专业知识。

An example of what we’re doing is a product we created that’s called AutoML. This is a unique product on the market to really lower the entry barrier of AI as much as much as possible—so that AI can be delivered to people who don’t do AI. The customer pain point is that so many businesses need customized models to help them to tackle their own problems. So, in the context of computer vision, if say I were a retailer, I might need a model to recognize my logo. If I were National Geographic magazine, I might need a model to recognize wild animals. If I worked in the agricultural industry, I might need a model to recognize apples. People have all kinds of use cases, but not everybody has the machinery expertise to create the AI.

看到这个问题，我们开发了 AutoML 产品，这样只要有人知道他们需要什么，比如“我需要它来区分苹果和橘子”，并且你提供训练数据，我们就会为你做所有事情。所以，从你的角度来看，这一切都是自动的，并为你的问题提供定制的机器学习模型。我们在一月份推出了 AutoML，成千上万的客户已经注册了这项服务。看到这种尖端人工智能的民主化，我感到非常欣慰。

Seeing this problem, we built the AutoML product so that as long as someone knows what they need, such as, “I need it for apples versus oranges,” and you bring the training data, we will do everything for you. So, from your perspective, it’s all automatic and delivers a customized machine learning model for your problem. We rolled AutoML out in January, and tens of thousands of customers have signed up to this service. It’s been very rewarding to see this democratization of cutting-edge AI.

马丁·福特：听起来，如果 AutoML 可以让技术水平较低的人也能使用机器学习，那么很容易导致由不同人为不同目标创建的各种人工智能应用程序的爆炸式增长。

MARTIN FORD: It sounds like AutoML, if it makes machine learning accessible to less technical people, could easily result in a sort of explosion of all kinds of AI applications created by different people with different objectives.

李飞飞：没错！事实上，我在一次演讲中用了寒武纪生命大爆发的比喻。

FEI-FEI LI: Yes, exactly! In fact, I used a Cambrian explosion analogy in one of my presentations.

马丁·福特：如今，人们非常关注神经网络和深度学习。您认为这是未来的发展方向吗？您显然相信深度学习会随着时间的推移而不断完善，但您认为它真的是引领人工智能走向未来的基础技术吗？或者是否存在另一种完全不同的东西，我们最终会抛弃深度学习和反向传播等，而采用一些全新的东西？

MARTIN FORD: Today there is an enormous focus on neural networks and deep learning. Do you think that’s the way forward? You obviously believe that deep learning will be refined over time, but you do think that it is really the foundational technology that’s going to lead AI into the future? Or is there another thing out there that’s completely different, where we’re going to end up throwing away deep learning and back propagation and all of that, and have something entirely new?

李飞飞：纵观人类文明，科学进步的道路总是建立在自我毁灭的基础上。历史上从未有过科学家说没有未来可言，没有改进可言。对于人工智能来说尤其如此，这是一个刚刚诞生 16 年的新兴领域。与拥有数百年甚至数千年历史的物理、生物和化学等领域相比，人工智能仍有许多进步空间。

FEI-FEI LI: If you look at human civilization, the path of scientific progress is always built upon undoing yourself. There isn’t a moment in history where scientists would have said that there’s nothing more to come, that there’s no refinement left. This is especially true for AI, which is such a nascent field that’s only been around for 16 years. Compared to fields like physics, biology, and chemistry, which have hundreds if not thousands of years of history, AI still has a lot to progress.

作为一名人工智能科学家，我从哲学角度并不认为我们已经完成了任务，也不认为卷积神经网络和深度学习是解决一切问题的答案——远远不够。正如你之前所说，很多问题都不是标记数据，也不涉及大量训练示例。回顾文明史及其教给我们的东西，我们不可能认为我们已经到达了目的地。正如我两岁大的孩子逃离婴儿床的故事告诉我们的那样，我们没有任何人工智能能够接近这种智能水平。

As an AI scientist, I do not philosophically believe we’ve finished our task, that convolutional neural networks and deep learning are the answers to everything—not by a huge margin. As you said earlier, a lot of problems are not labeled data or involve lots of training examples. Looking at the history of civilization and the things it’s taught us, we cannot possibly think we’ve reached a destination yet. As my two-year-old kid escaping the crib story tells us, we don’t have any AI that is close to that level of intelligence sophistication.

马丁·福特：您认为哪些具体项目处于人工智能研究的前沿？

MARTIN FORD: What particular projects would you point to that you think are at the forefront of research in AI?

李飞飞：在我的实验室里，我们一直在做一个远远超越 ImageNet 的项目，叫做“视觉基因组计划”。在这个项目中，我们深入思考了视觉世界，我们认识到 ImageNet 非常贫乏。ImageNet 只是对图片或视觉场景中的物体给出了一些离散的标签，而在真实的视觉场景中，物体是相互联系的，人类和物体在做很多事情。视觉和语言之间也有联系，所以视觉基因组计划实际上是人们所说的超越 ImageNet 的下一步。它旨在真正关注视觉世界和我们的语言之间的关系，所以我们一直在做大量工作来推动它的发展。

FEI-FEI LI: In my own lab, we have been doing a project that goes way beyond ImageNet, called the Visual Genome Project. In this project, we’ve thought deeply about the visual world, and we have recognized that ImageNet is very impoverished. ImageNet just gives some discreet labels of objects on the picture or visual scene, whereas in real visual scenes, objects are connected, humans and objects are doing a lot of things. There’s also the connection between vision and language, so Visual Genome Project is really what one would call the next step beyond ImageNet. It’s designed to really focus on the relationships between the visual world and our language, so we’ve been doing a lot of work in pushing that forwards.

另一个让我非常兴奋的方向是人工智能和医疗保健。我们实验室目前正在开展一个项目，该项目的灵感来自对医疗保健的一个特定元素的关注——护理本身。护理这个话题与很多人有关。护理是照顾病人的过程，但如果你看看我们的医院系统，就会发现有很多效率低下的地方：护理质量低下、缺乏监控、错误多多，整个医疗保健流程的成本高昂。看看外科手术中的错误，以及可能导致医院感染的卫生条件差。老年家庭护理也缺乏帮助和意识。护理方面存在很多问题。

Another direction I’m super excited about involves AI and healthcare. We’re currently working on a project in my lab that’s inspired by a focus on one particular element of healthcare—care itself. The topic of care touches a lot of people. Care is the process of taking care of patients, but if you look at our hospital system, for example, there are a lot of inefficiencies: low quality care, lack of monitoring, errors, and high costs associated with the whole healthcare delivery process. Just look at the mistakes of the surgery world, and the lack of hygiene that can result in hospitals acquiring infections. There’s also the lack of help and of awareness in senior home care. There are a lot of problems in care.

大约五年前，我们就意识到，有助于医疗服务的技术与自动驾驶汽车和人工智能等顶尖技术非常相似。我们需要智能传感器来感知环境和情绪，我们需要算法来理解收集到的数据并向临床医生、护士、患者和家属提供反馈。因此，我们开始开拓医疗服务领域人工智能的研究。我们正在与斯坦福儿童医院、犹他州山间医院和旧金山的开放式养老院合作。我们最近在《新英格兰医学杂志》上发表了一篇评论文章。我认为这非常令人兴奋，因为它使用了尖端的人工智能技术，就像自动驾驶汽车使用的技术一样，但它被应用于对人类需求和福祉至关重要的领域。

We recognized, about five years ago, that the technology that could help healthcare delivery is very similar to the top technology of self-driving cars and AI. We need smart sensors to sense the environment and the mood, and we need algorithms to make sense of the data collected and give feedback to clinicians, nurses, the patient and family members. So, we started pioneering this AI for healthcare delivery area of research. We’re working with Stanford children’s hospital, Utah’s Intermountain Hospital, and San Francisco’s unlocked senior homes. We recently published an opinion piece in the New England Journal of Medicine. I think it’s very exciting because it’s using cutting-edge AI technology, like the ones self-driving cars use, but it’s applied to an area that is so deeply critical for human needs and wellbeing.

马丁·福特：我想谈谈通用人工智能（AGI）的发展道路。您认为我们需要克服的主要障碍是什么？

MARTIN FORD: I want to talk about the path to artificial general intelligence (AGI). What you think the major hurdles that we would need to surmount are?

李飞飞：我想分两部分来回答你的问题。第一部分我会更狭义地回答关于 AGI 之路的问题；第二部分，我想谈谈我认为未来 AI 发展的框架和思路应该是什么样的。

FEI-FEI LI: I want to answer your question in two parts. The first part I’ll answer more narrowly on the question about the path to AGI, and in the second part, I want to talk about what I think the framework and frame of mind should be for the future development of AI.

因此，我们首先来定义一下 AGI，因为这不是 AI 与 AGI 的对决：它们都处于一个连续体上。我们都认识到，当今的 AI 非常狭窄且任务特定，专注于使用标记数据进行模式识别，但随着我们让 AI 变得更加先进，这一界限将变得宽松，因此在某种程度上，AI 和 AGI 的未来是一个模糊的定义。我猜 AGI 的一般定义应该是情境化、情境感知、细致入微、多方面和多维度的智能——并且具有人类的学习能力，这种学习能力不仅通过大数据，还通过无监督学习、强化学习、虚拟学习和各种学习来实现。

So, let’s first define AGI, because this isn’t about AI versus AGI: it’s all on one continuum. We all recognize today’s AI is very narrow and task specific, focusing on pattern recognition with labeled data, but as we make AI more advanced, that is going to be relaxed, and so in a way, the future of AI and AGI is one blurred definition. I guess the general definition of AGI would be the kind of intelligence that is contextualized, situationally aware, nuanced, multifaceted and multidimensional—and one that has the kind of learning capability that humans do, which is not only through big data but also through unsupervised learning, reinforcement learning, virtual learning, and various kinds of learning.

如果我们以此作为 AGI 的定义，那么我认为 AGI 之路就是不断探索超越监督的算法。我还认为，认识到脑科学、认知科学和行为科学的跨学科合作需求也很重要。许多人工智能技术，无论是任务的假设，还是算法的评估或猜想，都可以涉及脑科学和认知科学等相关领域。投资和倡导这种合作和跨学科方法也至关重要。我实际上在 2018 年 3 月的《纽约时报》社论中写过这篇文章，标题为《如何让人工智能造福人类》。

If we use that as a definition of AGI, then I think the path to AGI is a continued exploration of algorithms that are beyond just supervised. I also believe that it’s important to recognize the interdisciplinary need for collaborations in brain science, cognitive science, and behavior science. A lot of AI’s technology, whether it’s the hypothesis of the task or the evaluation or the conjecture of algorithms, can touch on related areas like brain science and cognitive science. It’s also really critical that we invest and advocate for this collaboration and interdisciplinary approach. I’ve actually written about this in my New York Times editorial opinion piece in March 2018 titled, How to Make A.I. That’s Good for People.

马丁·福特：是的，我读到了这篇文章，我知道您一直在倡导为人工智能发展的下一阶段建立一个全面的框架。

MARTIN FORD: Right, I read that, and I know you’ve been advocating a comprehensive framework for the next phase of AI development.

李飞飞：是的，我之所以这么做，是因为人工智能已经从一个学术和小众学科领域发展成为一个更大的领域，对人类生活产生了深远的影响。那么，我们如何划分人工智能，我们如何在下一阶段为未来创造人工智能？

FEI-FEI LI: Yes, I did that because AI has graduated from an academic and niche subject area into a much bigger field that impacts human lives in very profound ways. So how do we divide AI, and how do we create AI in the next phase, for the future?

以人为本的人工智能有三个核心组成部分或要素。第一个组成部分是推动人工智能本身的发展，这与我刚才谈到的跨学科研究和工作有很大关系，包括人工智能、神经科学和认知科学。

There are three core components or elements of human-centered AI. The first component is advancing AI itself, which has a lot to do with what I was just talking about: interdisciplinary research and work, on AI, across neuroscience and cognitive science.

以人为本的人工智能的第二个组成部分实际上是技术和应用；以人为本的技术。我们经常谈论人工智能在工作场景中取代人类，但人工智能在增强人类和增强人类方面还有更多机会。机会要多得多，我认为我们应该倡导和投资于人机协作和互动的技术。这就是机器人技术、自然语言处理、以人为本的设计等等。

The second component of human-centered AI is really the technology and application; the human-centered technology. We talk a lot about AI replacing humans in terms of a job scenario, but there are way more opportunities for AI to enhance humans and augment humans. The opportunities are much, much wider and I think we should advocate and invest in technology that is about collaboration and interaction between humans and machines. That’s robotics, natural language processing, human-centric design, and all that.

以人为本的人工智能的第三个组成部分是认识到，仅靠计算机科学无法解决所有人工智能机遇和问题。人工智能是一项对人类影响深远的技术，因此我们应该让经济学家来谈论就业、更大的组织和金融。我们应该让政策制定者、法律学者和伦理学家来谈论法规、偏见、安全和隐私。我们应该与历史学家、艺术家、人类学家和哲学家合作，研究人工智能研究的不同含义和新领域。这些实际上是下一阶段以人为本的人工智能的三个要素。

The third component of human-centered AI recognizes that computer science alone cannot address all the AI opportunities and issues. It’s a deeply impactful technology to humanity, so we should be bringing in economists to talk about jobs, to talk about bigger organizations, to talk about finance. We should bring in policymakers and law scholars and ethicists to talk about regulations, to talk about bias, to talk about security and privacy. We should work with historians, with artists, with anthropologists, and with philosophers—to look at the different implications and new areas of AI research. These are really the three elements that are about human-centered AI for the next phase.

马丁·福特：当你谈到以人为本的人工智能时，你试图解决一些已经提出的问题，我想谈谈其中的一些。有一种观点认为，存在着真正的生存威胁，这是尼克·博斯特罗姆、埃隆·马斯克和斯蒂芬·霍金提出的观点，超级智能可能会非常迅速地出现，这是一个递归的自我完善循环。我听人说，你的 AutoML 可能是朝着这个方向迈出的一步，因为你正在使用技术来设计其他机器学习系统。你对此有什么看法？

MARTIN FORD: When you talk about human-centered AI, you’re trying to address some concerns that have been raised, and I wanted to touch on some of those. There’s this idea that there is a true existential threat, something that’s been raised by Nick Bostrom, Elon Musk, and Stephen Hawking, where super intelligence could happen very rapidly, a recursive self-improvement loop. I’ve heard people say that your AutoML might be one step toward that because you’re using technology to design other machine learning systems. What do you think about that?

李飞飞：我认为，我们有像尼克·博斯特罗姆这样的思想领袖来推测人工智能的未来相当令人不安，或者至少发出警告信号，警告我们可能会以我们意想不到的方式影响我们，这是很正常的。但我认为，重要的是要将其置于背景中，因为在人类文明的悠久历史中，每当一种新的社会秩序或技术被发明时，它都有可能以意想不到的、深刻的方式颠覆人类世界。

FEI-FEI LI: I think that it’s healthy that we have thought leaders like Nick Bostrom to conjecture a fairly troubling future of AI, or at least send warning signs of things that could impact us in ways that we didn’t expect. But I think it’s important to contextualize that, because in the long history of human civilization, every time a new social order or technology has been invented, it’s had that same potential to disrupt the human world in unexpected and deeply profound ways.

我还认为，通过多种多样的声音来探索这些重要问题是有益的。这些声音来自不同的发展路径。有哲学家尼克来哲学化这些潜力是件好事。尼克的声音是人工智能社会话语中的一种。我认为我们需要多种声音来做出贡献。

I also think that it’s healthy to have different ways of exploring these important questions through a diversity of voices. And from voices coming from different development paths. It’s good to have Nick, who is a philosopher, to philosophize the potentials. Nick’s is one type of voice in the social discourse of AI. I think we need many voices to contribute.

马丁·福特：伊隆·马斯克等人确实非常重视这种担忧，他曾说人工智能比朝鲜的威胁更大，引起了广泛关注。你认为这种说法是否有些过分？或者，我们整个社会现在真的应该如此担忧吗？

MARTIN FORD: That particular concern really has been given a lot of weight by people like Elon Musk who attracts a lot of attention by saying, for example, that AI is a bigger threat than North Korea. Do you think that’s over the top, or should we really be that concerned as a society, at this point in time?

李飞飞：从定义上讲，我们倾向于记住那些夸张的陈述。作为一名科学家和学者，我倾向于关注那些建立在更深层次、有充分证据和逻辑推理基础上的论点。我是否评判某个句子其实并不重要。

FEI-FEI LI: By definition, we tend to remember over-the-top statements. As a scientist and as a scholar I tend to focus on arguments that are built on deeper and well substantiated evidence and logical deduction. It’s really not important whether I judge a particular sentence or not.

重要的是我们如何利用现在的机会，以及我们每个人正在做什么。例如，我更愿意谈论人工智能中的偏见和缺乏多样性，这就是我所谈论的，因为看看我做了什么更重要。

The important thing is what we do with the opportunities we have now, and what each one of us is doing. For example, I’m more vocal to discuss the bias and lack of diversity in AI, and so that’s what I speak about, because it’s much more important to look at what I do.

马丁·福特：那么，生存威胁还远在未来吗？

MARTIN FORD: So, the existential threat is pretty far in the future?

李飞飞：就像我说的，有些人开始思考生存威胁，这是很正常的。

FEI-FEI LI: Well like I said, it’s healthy that some people are thinking about that existential threat.

马丁·福特：您简要提到了对就业的影响，这是我写过很多次的话题，事实上，这也是我上一本书的主题。您说，肯定存在提升人才的机会，但与此同时，技术与资本主义之间存在着交集，如果可以的话，企业总是有非常强烈的动机来消除劳动力。历史上一直都是这样。今天我们似乎正处于一个转折点，很快就会出现能够自动化范围比过去更广泛的任务的工具。这些工具将取代认知和智力任务，而不仅仅是体力劳动。是否会出现大量失业、工作技能下降、工资下降等现象？

MARTIN FORD: You mentioned briefly the impact on jobs, and this is something that I’ve written about a lot, in fact, it’s what my previous book was about. You said that there are definitely opportunities to enhance people—but at the same time, there is this intersection of technology and capitalism, and businesses always have a very strong motive to eliminate labor if they can. That’s happened throughout history. It seems as though we’re at an inflection point today, where there are soon going to be tools that are able to automate a much broader range of tasks than anything in the past. These tools will replace cognitive and intellectual tasks, and not just manual work. Is there potential for lots of job losses, deskilling of jobs, depressed wages, and so forth?

李飞飞：我并不想成为一名经济学家，但资本主义是人类社会秩序的一种形式，它已经有 100 年的历史了？我想说的是，没有人能够预测资本主义是人类社会向前发展的唯一形式；也没有人能够预测未来社会的技术将如何演变。

FEI-FEI LI: I don’t pretend to be an economist, but capitalism is one form of human societal order and it is what, 100 years old? What I’m saying is that no one can predict that capitalism is the only form of human society going forward; nor can anyone predict how technology is going to morph in that future society.

我的观点是，人工智能作为一种具有巨大潜力的技术，有机会让生活变得更好，让工作更有效率。我和医生一起工作了五年，我知道医生的某些工作可能会被机器取代。但我真的希望这部分工作能被取代，因为我看到我们的医生工作过度，不堪重负，他们的才华有时没有得到应有的发挥。我希望我们的医生有时间与患者交谈，有时间相互交流，有时间了解和优化疾病的最佳治疗方法。我希望我们的医生有时间做一些罕见或更难治的疾病所需的侦查工作。

My argument is that AI, as a technology with a lot of potentials, has an opportunity to make life a lot better, to make work more productive. I’ve been working with doctors for five years, and I get that there are parts of doctors’ work that can be potentially replaced by a machine. But I really want that part to be replaced, because I see our doctors overworked, overwhelmed, and their brilliance is sometimes not used in the ways that it should be used. I want to see our doctors having time to talk to patients, having time to talking to each other, and having time to understand and optimize for the best treatment of diseases. I want to see our doctors having time to do the detective work that some rare or harder illnesses need.

人工智能作为一种技术，除了取代劳动力之外，还有很大的潜力来增强和增加劳动力，我希望我们能看到越来越多的这种技术。这是我们在历史上看到的。大约 40 年前，计算机使许多工作从办公室打字员手中脱离出来。但我们看到的是新的工作，我们现在有了软件工程师这一新工作，人们在办公室里做着更有趣的工作。自动取款机也是如此：当他们开始在银行自动化某些交易时，柜员的数量实际上增加了，因为有更多的金融服务可以由人类完成——而普通的现金存取现在可以由自动取款机完成。这根本不是一个黑白分明的故事，是我们共同的努力决定了事情的发展方向。

AI as a technology has so much potential to enhance and augment labor, in addition to just replace it, and I hope that we see more and more of that. This is something that we’ve got evidence of in history. Computers automated a lot of jobs away from office typists some 40 years ago. But what we see is new jobs, we now have software engineers as a new job, we have people doing way more interesting work around the office. The same went with ATM machines: when they started to automate some of these transactions in the bank, the number of tellers actually increased because there were more financial services that could be done by humans—and the mundane cash deposit or withdrawal can now be done by ATM machines. It’s not a black and white story at all, and it’s our work together that defines how things go.

马丁·福特：我们来谈谈您关注的其他一些话题，比如多样性和偏见。在我看来，这实际上是两个独立的事情。偏见的产生是因为它包含在机器学习算法所训练的人类生成的数据中，而多样性则更多地取决于谁在从事人工智能工作。

MARTIN FORD: Let’s talk about some of the other topics you’ve focused on, like diversity and bias. It seems to me that these are two separate things, really. The bias comes about because it’s encapsulated in the human-generated data that machine learning algorithms are trained on, whereas diversity is more of an issue of who’s working in AI.

李飞飞：首先，我认为它们并不像你想象的那么独立，因为归根结底，它们都是人类赋予机器的价值。如果我们有一个机器学习管道，从数据本身开始，那么当数据有偏差时，我们的机器学习结果也会有偏差。某些形式的偏差甚至可能带来致命的影响。但这本身可能与管道的开发过程有关。我只是想从哲学角度提出一个观点，即它们实际上可能存在联系。

FEI-FEI LI: First of all, I don’t think they’re as separate as you think because at the end of the day, it’s the values that humans bring to machines. If we have a machine learning pipeline, starting with the data itself, then when that data is biased our machine learning outcome will be biased. And some forms of bias might have even fatal implications. But that itself is potentially linked to the development process of the pipeline. I just want to make a philosophical point that they are actually potentially linked.

话虽如此，我同意你的观点，偏见和多样性可以分开处理。例如，就数据偏见导致机器学习结果偏见而言，许多学术研究人员现在已经认识到这一点，并正在努力寻找揭露这种偏见的方法。他们还在修改算法以应对偏见，并试图通过这种方式纠正它。从学术界到工业界，这种对产品和技术偏见的暴露确实很健康，它让整个行业保持警惕。

That now said, I agree with you that bias and diversity can be treated a little more separately. For example, in terms of data bias resulting in machine learning outcome bias, a lot of academia researchers are recognizing this now, and working on ways to expose that kind of bias. They’re also modifying algorithms to respond to bias in a way to try to correct it that way. This exposure to the bias of products and technology, from academia to industry, is really healthy, and it keeps the industry on their toes.

马丁·福特：你肯定要处理谷歌的机器学习偏见问题。你是如何解决这个问题的？

MARTIN FORD: You must have to deal with machine learning bias at Google. How do you address it?

李飞飞：谷歌现在有一大批研究人员致力于研究机器学习偏见和“可解释性”，因为压力在于解决偏见问题、提供更好的产品，我们希望帮助他人。现在还处于早期阶段，但投资这一研究领域并取得更多进展至关重要。

FEI-FEI LI: Google now has a whole group of researchers working on machine learning bias and “explainability” because the pressure is there to tackle bias, to deliver a better product, and we want to be helping others. It’s still early days, but it’s so critical that this area of research gets invested and that there’s more development in that.

关于多样性和人们的偏见，我认为这是一场巨大的危机。我们还没有解决劳动力多样性的问题，尤其是在 STEM 领域。然后，由于技术和人工智能还处于萌芽阶段，但作为一种技术却具有如此大的影响力，这个问题更加严重。如果我们环顾四周，无论是公司里的人工智能团队、学术界的人工智能教授、人工智能博士生还是顶级人工智能会议上的人工智能演讲者，无论你从哪个角度看：我们缺乏多样性。我们缺乏女性，我们缺乏代表性不足的少数群体。

On the topic of diversity and the bias of people, I think it’s a huge crisis. We’ve not solved the issue of diversity in our workforces, especially in STEM. Then with tech and AI being so nascent and yet so impactful as a technology, this problem is exacerbated. If we look around, whether you’re looking at AI groups in companies, AI professors in academia, AI PhD students or AI presenters at top AI conferences, no matter where you cut it: we lack diversity. We lack women, and we lack under-represented minorities.

马丁·福特：我知道你启动了 AI4ALL 项目，该项目致力于吸引女性和代表性不足的少数群体进入人工智能领域。你能谈谈这个项目吗？

MARTIN FORD: I know you started the AI4ALL project, which is focused on attracting women and underrepresented minorities into the field of AI. Could you talk about that?

李飞飞：是的，我们一直在讨论的代表性不足问题促使我在四年前启动了斯坦福 AI4ALL 项目。我们可以做出的一项重要努力是在高中生上大学并决定专业和未来职业之前激励他们，并邀请他们参与人工智能研究和学习。我们尤其认为，对于那些受到人工智能人类使命启发的代表性不足的少数群体来说，他们会对那种超越自我的动机和灵感做出回应。因此，过去四年来，我们每年都在斯坦福大学制定这个暑期课程，并邀请高中女生参与人工智能。这个项目非常成功，因此我们在 2017 年成立了一个名为 AI4ALL 的全国性非营利组织，并开始复制这种模式并邀请其他大学参与。

FEI-FEI LI: Yes, that lack of representation we’ve been discussing led to me start the Stanford AI4ALL project four years ago. One important effort we can make is to inspire high school students, before they go to college and decide on their major and future career, and to invite them into AI research and AI study. We especially think that, for underrepresented minorities who are inspired by human missions in AI, they respond to the kind of bigger-than-themselves motivations and inspiration. As a result, we’ve crafted this summer curriculum every year at Stanford, for the past four years, and invited high school girls to participate in AI. This was so successful that in 2017 we formed a national nonprofit organization called AI4ALL and started to replicate this model and invite other universities to participate.

一年后，我们有六所大学针对人工智能领域，而这些领域很难吸引人们参与。除了斯坦福大学和西蒙弗雷泽大学，伯克利大学也针对低收入学生开展人工智能，普林斯顿大学专注于为少数族裔开展人工智能，克里斯托弗纽波特大学为非正规学生开展人工智能，波士顿大学为女孩开展人工智能。这些项目只运行了一小段时间，但我们希望扩大项目规模，并继续邀请来自更多样化背景的未来人工智能领袖。

A year later, we have six universities targeting different areas where AI has really struggled to get people involved. In addition to Stanford and Simon Fraser University, we’ve also got Berkeley targeting AI for low-income students, Princeton focusing on AI for racial minorities, Christopher Newport University doing AI for off-the-rails students, and Boston University doing AI for girls. These have only been running for a small amount of time, but we’re hoping to mushroom the program and continue to invite future leaders of AI from a much more diverse background.

马丁·福特：我想问一下，您是否认为人工智能需要监管。这是您希望看到的吗？您是否主张政府在制定规则方面投入更多精力，或者您认为人工智能社区可以在内部解决这些问题？

MARTIN FORD: I wanted to ask if you think there’s a place for the regulation of artificial intelligence. Is that something you’d like to see? Would you advocate for the government taking more of an interest, in terms of making rules, or do you think that the AI community can solve these problems internally?

李飞飞：事实上，我不认为人工智能（如果你指的是人工智能技术人员）可以自己解决所有人工智能问题：我们的世界是相互联系的，人类的生活是相互交织的，我们都相互依赖。

FEI-FEI LI: I actually don’t think AI, if you mean the AI technologists, can solve all the AI problems by themselves: our world is interconnected, human lives are intertwined, and we all depend on each other.

无论我实现了多少人工智能，我仍然在同一条高速公路上行驶，呼吸着同样的空气，送我的孩子去社区学校。我认为我们需要以非常人性化的方式看待这个问题，并认识到，任何技术要想产生如此深远的影响，我们都需要邀请生活和社会的各个领域参与其中。

No matter how much AI that I make happen, I still drive on the same highway, breathe the same air, and send my kids to community schools. I think that we need to have a very humanistic view of this and recognize that for any technology to have this profound impact, we need to invite all sectors of life and society to participate.

我还认为政府发挥着巨大的作用，即投资人工智能的基础科学、研究和教育。因为如果我们想要拥有透明的技术，如果我们想要拥有公平的技术，如果我们想要让更多人能够理解并以积极的方式影响这项技术，那么政府就需要投资我们的大学、研究机构和学校，教育人们了解人工智能并支持基础科学研究。我没有接受过政策制定者的培训，但我与一些政策制定者交谈过，我也与我的朋友交谈过。无论是关于隐私、公平、传播还是合作，我都看到政府可以发挥的作用。

I also think the government has a huge role, which is to invest in basic science, research, and education of AI. Because if we want to have the transparent technology, and if we want to have the fair technology, and if we want to have more people who can understand and impact this technology in positive ways, then the government needs to invest in our universities, research institutes and schools to educate people about AI and support basic science research. I’m not trained as a policymaker, but I talk to some policymakers, and I talk to my friends. Whether it’s about privacy, fairness, dissemination, or collaboration, I see a role the government can play.

马丁·福特：我最后想问您的是这场人工智能军备竞赛，尤其是与中国的军备竞赛。您对此有多重视？这是我们应该担心的事情吗？

MARTIN FORD: The final thing I want to ask you about is this perceived AI arms race, especially with China. How seriously do you take that, and is it something we should worry about?

中国确实有一个不同的体制，一个更专制的体制，以及更大的人口，这意味着有更多的数据来训练算法，而隐私等方面的限制更少。我们是否有落后于人工智能领导地位的风险？

China does have a different system, a more authoritarian system, and a much bigger population which means more data to train algorithms on and less restrictions regarding privacy and so forth. Are we at risk of falling behind in AI leadership?

李飞飞：现在，我们正处于现代物理学的大爆炸周期中，它正在改变技术，无论是核技术还是电气技术。

FEI-FEI LI: Right now, we’re living in a major hype-cycle of modern physics and how that can transform technology, whether it’s nuclear technology, or electrical technology.

一百年后，我们会问自己这个问题：谁拥有现代物理学？我们会试图说出拥有现代物理学和工业革命后一切的公司或国家吗？我认为我们任何人都很难回答这些问题。作为一名科学家和教育工作者，我的观点是，人类对知识和真理的追求是没有国界的。如果科学有一个基本原则，那就是这些都是普遍的真理和对这些真理的​​追求，我们作为一个物种共同寻求这些真理。在我看来，人工智能是一门科学。

One hundred years later, will we ask ourselves the question: which person owned modern physics? Will we try to name the company or country that owned modern physics and everything after the industrial revolution? I think it will be difficult for any of us to answer those questions. My point is, as a scientist and as an educator, that the human quest for knowledge and truth has no borders. If there is a fundamental principle of science, it is that these are the universal truths and quests for these truths, which we all seek as a species together. And AI is a science in my opinion.

从这个角度来看，作为一名基础科学家和教育工作者，我与来自不同背景的人一起工作。我的斯坦福实验室实际上由来自各大洲的学生组成。我们希望我们创造的技术，无论是自动化还是医疗保健，都能造福所有人。

From that point of view, as a basic scientist and as an educator, I work with people from all backgrounds. My Stanford lab literally consists of students from every continent. With the technology we create, whether it’s automation or it’s healthcare, we hope to benefit everyone.

当然，公司之间和地区之间会存在竞争，我希望这种竞争是良性的。良性竞争意味着我们相互尊重，尊重市场，尊重用户和消费者，尊重法律，即使是跨境法律或国际法。作为一名科学家，这就是我所提倡的，我将继续在开源领域发表文章，教育各种肤色和国家的学生，我希望与各种背景的人合作。

Of course, there is going to be competition between companies and between regions, and I hope that’s healthy. Healthy competition means that we respect each other, we respect the market, we respect the users and consumers, and we respect the laws, even if it’s cross-border laws or international laws. As a scientist, that’s what I advocate for, and I continue to publish in the open source domain to educate students of all colors and nations, and I want to collaborate with people of all backgrounds.

有关 AI4ALL 的更多信息，请访问 http://ai-4-all.org/ 。

More Information about AI4ALL can be found at http://ai-4-all.org/ .

李飞飞 是 Google Cloud 人工智能和机器学习首席科学家、斯坦福大学计算机科学教授，也是斯坦福人工智能实验室和斯坦福视觉实验室的主任。李飞飞在普林斯顿大学获得物理学学士学位，在加州理工学院获得电气工程博士学位。她的工作重点是计算机视觉和认知神经科学，并在顶级学术期刊上发表了大量文章。她是 AI4ALL 的联合创始人，该组织致力于吸引女性和来自代表性不足群体的人进入人工智能领域，该组织始于斯坦福大学，现已扩展到美国各地的大学。

FEI-FEI LI is Chief Scientist, AI and Machine Learning at Google Cloud, Professor of Computer Science at Stanford University, and Director of both the Stanford Artificial Intelligence Lab and the Stanford Vision Lab. Fei-Fei received her undergraduate degree in physics from Princeton University and her PhD in electrical engineering from the California Institute of Technology. Her work has focused on computer vision and cognitive neural science and she is widely published in top academic journals. She is the co-founder of AI4ALL, an organization focused on attracting women and people from underrepresented groups into the field of AI, which began at Stanford and has now scaled up to universities across the United States.

DEEPMIND 联合创始人兼首席执行官 人工智能研究员和神经科学家

CO-FOUNDER & CEO OF DEEPMIND AI RESEARCHER AND NEUROSCIENTIST

德米斯·哈萨比斯曾是一名儿童国际象棋神童，16 岁开始专业地编写和设计电子游戏。从剑桥大学毕业后，德米斯花了十年时间领导和创立了专注于电子游戏和模拟的成功初创公司。他回到学术界，在伦敦大学学院攻读认知神经科学博士学位，随后在麻省理工学院和哈佛大学从事博士后研究。他于 2010 年与他人共同创立了 DeepMind。DeepMind 于 2014 年被谷歌收购，现在是 Alphabet 旗下公司的一部分。

Demis Hassabis is a former child chess prodigy, who started coding and designing video games professionally at age 16. After graduating from Cambridge University, Demis spent a decade leading and founding successful startups focused on video games and simulation. He returned to academia to complete a PhD in cognitive neuroscience at University College London, followed by postdoctoral research at MIT and Harvard. He co-founded DeepMind in 2010. DeepMind was acquired by Google in 2014 and is now part of Alphabet’s portfolio of companies.

马丁·福特：我知道你年轻时对国际象棋和电子游戏非常感兴趣。这对你的人工智能研究生涯以及你创立 DeepMind 的决定有何影响？

MARTIN FORD: I know you had a very strong interest in chess and video games when you were younger. How has that influenced your career in AI research and your decision to found DeepMind?

德米斯·哈萨比斯：我小时候是一名职业国际象棋选手，梦想成为国际象棋世界冠军。我是一个内省的孩子，我想提高自己的棋艺，所以我经常思考我的大脑是如何想出这些走法的。当你走好一步或犯错时，大脑中发生了哪些过程？所以，我很早就开始思考思考的问题，这导致我在后来的生活中对神经科学等事物产生了兴趣。

DEMIS HASSABIS: I was a professional chess player in my childhood with aspirations of becoming the world chess champion. I was an introspective kid and I wanted to improve my game, so I used to think a lot about how my brain was coming up with these ideas for moves. What are the processes that are going on there when you make a great move or a blunder? So, very early on I started to think a lot about thinking, and that led me to my interest in things like neuroscience later on in my life.

当然，国际象棋在人工智能中扮演着更重要的角色。自人工智能诞生以来，国际象棋本身一直是人工智能研究的主要问题领域之一。艾伦·图灵和克劳德·香农等人工智能领域的一些早期先驱对计算机国际象棋非常感兴趣。我 8 岁时，用参加国际象棋比赛的奖金购买了我的第一台电脑。我记得我编写的第一个程序是奥赛罗游戏（也称为黑白棋），虽然它比国际象棋简单，但我使用了早期人工智能先驱在他们的国际象棋程序中使用的相同想法，例如 alpha-beta 搜索等。那是我第一次接触编写人工智能程序。

Chess, of course, has a deeper role in AI. The game itself has been one of the main problem areas for AI research since the dawn of AI. Some of the early pioneers in AI like Alan Turing and Claude Shannon were very interested in computer chess. When I was 8 years old, I purchased my first computer using the winnings from the chess tournaments that I entered. One of the first programs that I remember writing was for a game called Othello—also known as Reversi—and while it’s a simpler game than chess, I used the same ideas that those early AI pioneers had been using in their chess programs, like alpha-beta search, and so on. That was my first exposure to writing an AI program.

我对国际象棋和游戏的热爱让我开始编程，特别是为游戏编写人工智能。我的下一个阶段是将我对游戏和编程的热爱融入到商业电子游戏的编写中。在我的许多游戏中，从《主题公园》（1994 年）到《共和国：革命》（2003 年），你都会看到一个关键主题，那就是模拟是游戏玩法的核心。这些游戏为玩家提供了沙盒，沙盒中的角色会根据你的游戏方式做出反应。人工智能是这些角色的基础，而这一直是我专门研究的部分。

My love of chess and games got me into programming, and specifically into writing AI for games. The next stage for me was to combine my love of games and programming into writing commercial videogames. One key theme that you’ll see in a lot of my games, from Theme Park (1994) to Republic: The Revolution (2003), was that they had simulation at the heart of their gameplay. The games presented players with sandboxes with characters in them that reacted to the way that you played. It was AI underpinning those characters, and that was always the part that I worked on specifically.

我通过游戏做的另一件事是训练我的大脑掌握某些能力。例如，我认为孩子们在学校学习国际象棋是一件很棒的事情，因为它教会了他们解决问题、规划和各种其他元技能，我认为这些技能在其他领域很有用，可以转化为实际应用。回想起来，当我创办 DeepMind 并开始使用游戏作为我们 AI 系统的训练环境时，所有这些信息可能都在我的潜意识中。

The other thing that I was doing with games was training my mind on certain capabilities. For example, with chess, I think it’s a great thing for kids to learn at school because it teaches problem-solving, planning, and all sorts of other meta-skills that I think are then useful and translatable to other domains. Looking back, perhaps all of that information was in my subconscious when I started DeepMind and started using games as a training environment for our AI systems.

在创办 DeepMind 之前，我的最后一步是在剑桥大学攻读计算机科学本科课程。当时，也就是 21 世纪初，我觉得我们这个领域还没有足够的想法去尝试攀登 AGI 的珠穆朗玛峰。这促使我攻读神经科学博士学位，因为我觉得我们需要更好地理解大脑如何解决其中一些复杂的功能，这样我们才能从中受到启发，提出新的算法想法。我学到了很多关于记忆和想象力的知识——这些主题我们当时不知道，在某些情况下，仍然不知道如何让机器去做。所有这些不同的线索后来汇聚在一起，形成了 DeepMind。

The final step for me, before starting DeepMind, was taking undergraduate computer science course at Cambridge University. At the time, which was the early 2000s, I felt that as a field we didn’t have quite enough ideas to try and attempt to climb the Everest of AGI. This led me to my PhD in Neuroscience because I felt we needed a better understanding of how the brain solved some of these complex capabilities, so that we could be inspired by that to come up with new algorithmic ideas. I learned a lot about memory and imagination—topics that we didn’t at the time, and in some cases still don’t, know how to get machines to do. All those different strands then came together into DeepMind.

马丁·福特：那么，从一开始，您的重点就是机器智能，尤其是 AGI 吗？

MARTIN FORD: Your focus then, right from the beginning, has been on machine intelligence and especially AGI?

德米斯·哈萨比斯：没错。我从十几岁起就知道自己想以此为职业。这段旅程始于我的第一台电脑。我立刻意识到电脑是一种神奇的工具，因为大多数机器都能扩展你的身体能力，但这里有一台机器可以扩展你的智力。

DEMIS HASSABIS: Exactly. I’ve known I wanted to do this as a career since my early teens. That journey started with my first computer. I realized straight away that a computer was a magical tool because most machines extend your physical capability, but here was a machine that could extend your mental capabilities.

我仍然对这样的事实感到兴奋：你可以编写一个程序来解决一个科学问题，让它运行，然后睡觉，第二天早上醒来时它就解决了。这几乎就像把你的问题外包给机器。这让我想到人工智能是自然而然的下一步，甚至是最后一步，我们让机器本身变得更聪明，这样它们不仅可以执行你给它们的命令，而且它们实际上能够自己想出解决方案。

I still get excited by the fact that you can write a program to crunch a scientific problem, set it running, go off to sleep, and then when you wake up in the morning it’s solved it. It’s almost like outsourcing your problems to the machine. This led me to think of AI as the natural next step, or even the end step, where we get machines to be smarter in themselves so they’re not just executing what you’re giving them, but they’re actually able to come up with their own solutions.

我一直想研究能够自我学习的学习系统，并且我一直对什么是智能以及我们如何人工地重现这种现象的哲学思想感兴趣，这也是我创建 DeepMind 的原因。

I’ve always wanted to work on learning systems that learn for themselves, and I’ve always been interested in the philosophical idea of what is intelligence and how can we recreate that phenomena artificially, which is what led me to create DeepMind.

马丁·福特：目前纯 AGI 公司的例子并不多。原因之一是，这种公司实际上没有商业模式；短期内很难产生收入。DeepMind 是如何克服这一困难的？

MARTIN FORD: There aren’t many examples of pure AGI companies around. One reason is that there’s not really a business model for doing that; it’s hard to generate revenue in the short term. How did DeepMind overcome that?

德米斯·哈萨比斯：从一开始，我们就是一家 AGI 公司，我们对此非常清楚。我们从一开始就有解决智能问题的使命宣言。你可以想象，试图向标准风险投资家推销这一点相当困难。

DEMIS HASSABIS: From the beginning, we were an AGI company, and we were very clear about that. Our mission statement of solving intelligence was there from the beginning. As you can imagine, trying to pitch that to standard venture capitalists was quite hard.

我们的论点是，由于我们正在开发的是一种通用技术，如果你能够将其构建得足够强大、足够通用、足够强大，那么它应该会有数百种令人惊叹的应用。你将会被各种可能性和机会淹没，但你需要先进行大量的前期研究，而这些研究需要我们聚集一群非常有才华的人。我们认为这是可以接受的，因为世界上真正能从事这项工作的人很少，特别是如果你回想一下我们刚开始的 2009 年和 2010 年。你可能数不清有不到 100 人能为这类工作做出贡献。那么问题来了，我们能否展示出明确且可衡量的进展？

Our thesis was that because what we were building was a general-purpose technology, if you could build it powerfully enough, general enough, and capable enough, then there should be hundreds of amazing applications for it. You’d be inundated with incoming possibilities and opportunities, but you would require a large amount of upfront research first from a group of very talented people that we’d need to get together. We thought that was defensible because of the small number of people in the world that could actually work on this, especially if you think back to 2009 and 2010 when we first started out. You could probably count less than 100 people that could contribute to that type of work. Then there was the question of can we demonstrate clear and measurable progress?

拥有一个庞大而长期的研究目标的问题是，你的资助者如何确信你真的知道自己在说什么？对于一家典型的公司来说，你的衡量标准是你的产品和用户数量，这是很容易衡量的。像 DeepMind 这样的公司之所以如此罕见，是因为对于外部非专业人士（如风险投资家）来说，很难判断你的计划是否合理，或者你只是疯了。

The problem with having a large and long-term research goal is how do your funders get confidence that you actually know what you’re talking about? With a typical company, your metric is your product and the number of users, something that’s easily measurable. The reason why a company like DeepMind is so rare is that’s very hard for an external non-specialist, like a venture capitalist, to judge whether you’re making sense and your plan really is sensible, or whether you’re just crazy.

这条界线非常模糊，尤其是当你走得很远的时候，而在 2009 年和 2010 年，没有人谈论人工智能。人工智能当时并不是如今的热门话题。由于人工智能领域过去 30 年未能实现承诺，我很难获得最初的种子资金。我们有一些非常有力的假设来解释为什么会这样，这些假设就是我们建立 DeepMind 的基础。比如从神经科学中汲取灵感，神经科学在过去 10 年里极大地提高了我们对大脑的理解；开发学习系统而不是传统的专家系统；使用基准测试和模拟来快速开发和测试人工智能。我们致力于的一系列事情最终被证明是正确的，它们解释了为什么人工智能在过去几年没有得到改进。另一个非常有力的事情是，这些新技术需要大量的计算能力，而这些计算能力现在正以 GPU 的形式提供。

The line is very thin, especially when you’re going very far out, and in 2009 and 2010 no one was talking about AI. AI was not the hot topic that it is today. It was really difficult for me to get my initial seed funding because of the previous 30 years of failed promises in AI. We had some very strong hypotheses as to why that was, and those were the pillars that we were basing DeepMind on. Things like taking inspiration from neuroscience, which had massively improved our understanding of the brain in the last 10 years; doing learning systems not traditional expert systems; using benchmarking and simulations for the rapid development and testing of AI. There was a set of things that we committed to that turned out to be correct and were our explanations for why AI hadn’t improved in the previous years. Another very powerful thing was that these new techniques required a lot of computing power, which was now becoming available in the form of GPUs.

我们的论点对我们来说是有意义的，最终，我们成功说服了足够多的人，但这很难，因为我们当时处于一个非常怀疑、不流行的领域。即使在学术界，人工智能也是不受欢迎的。它被重新命名为“机器学习”，从事人工智能工作的人被认为是边缘分子。看到这一切变化得如此之快，真是令人惊讶。

Our thesis made sense to us, and in the end, we managed to convince enough people, but it was hard because we were operating at that point within a very skeptical, non-fashionable domain. Even in academia, AI was frowned upon. It had been rebranded “machine learning,” and people who worked on AI were considered to be fringe elements. It’s amazing to see how quickly all of that has changed.

马丁·福特：最终，你们获得了资金，可以作为一家独立公司生存下去。但后来你们决定让谷歌收购 DeepMind。你能告诉我收购背后的原因以及收购过程吗？

MARTIN FORD: Eventually you were able to secure the funding to be viable as an independent company. But then you decided to let Google acquire DeepMind. Can you tell me about the rationale behind the acquisition and how that happened?

DEMIS HASSABIS：值得注意的是，我们并没有出售的计划，部分原因是我们认为在 DeepMind 开始生产产品之前，没有大公司会理解我们的价值。说我们没有商业模式也不公平。我们有，只是还没有在执行方面走得很远。我们确实已经有一些很酷的技术，DQN（深度 Q 网络——我们的第一个通用学习模型），我们的 Atari 工作已经在 2013 年完成了。但后来，谷歌联合创始人拉里·佩奇通过我们的一些投资者听说了我们，2013 年，我突然收到了谷歌搜索和研究部门负责人艾伦·尤斯塔斯的电子邮件，说拉里听说过 DeepMind，他想聊聊。

DEMIS HASSABIS: It’s worth noting that we had no plans to sell, partly because we figured no big corporate would understand our value until DeepMind started producing products. It’s also not fair to say that we didn’t have a business model. We did, we just hadn’t gone very far down the line of executing it. We did already have some cool technology, DQN (deep Q-network—our first general-purpose learning model) and our Atari work had already been done by 2013. But then Larry Page, the Co-Founder of Google, heard about us through some of our investors and out of the blue in 2013 I received an email from Alan Eustace, who was running search and research at Google, saying that Larry’s heard of DeepMind and he’d like to have a chat.

那只是个开始，但这个过程花了很长时间，因为在与谷歌联手之前，我想确定很多事情。但最终，我确信，通过结合谷歌的优势和资源——他们的计算能力和组建更大团队的能力，我们将能够更快地完成我们的使命。这与钱无关，我们的投资者愿意增加资金让我们独立运作，但 DeepMind 一直致力于提供 AGI 并将其用于造福世界，而与谷歌联手有机会加速这一进程。

That was the start, but the process took a long time because there were a lot of things I wanted to be sure of before we joined forces with Google. But at the end of the day, I became convinced that by combining with Google’s strengths and resources—their computing power and their ability to construct a much bigger team, we would be able to execute on our mission much more quickly. It wasn’t to do with money, our investors were willing to increase funding to keep us going independently, but DeepMind has always been about delivering AGI and using it for the benefit of the world, and there was an opportunity with Google to accelerate that.

拉里和谷歌的员工和我一样对人工智能充满热情，他们明白我们所做的工作有多么重要。他们同意给予我们自主权，让我们自主决定研究路线图和文化，并让我们留在伦敦，这对我来说非常重要。最后，他们还同意设立一个有关我们技术的道德委员会，这很不寻常，但他们很有先见之明。

Larry and the people at Google were just as passionate about AI as I was, and they understood how important the work we would do would be. They agreed to give us autonomy as to our research roadmap and our culture, and also to staying in London, which was very important to me. Finally, they also agreed to have an ethics board concerning our technology, which was very unusual but very prescient of them.

马丁·福特：为什么你选择在伦敦，而不是硅谷？这是因为 Demis Hassabis 还是 DeepMind？

MARTIN FORD: Why did you choose to be in London, and not Silicon Valley? Is that a Demis Hassabis or a DeepMind thing?

DEMIS HASSABIS：其实两者都有。我是土生土长的伦敦人，我爱伦敦，但同时，我认为这是一种竞争优势，因为英国和欧洲在人工智能领域拥有剑桥和牛津等出色的大学。而且，当时英国或欧洲还没有真正雄心勃勃的研究公司，所以我们的招聘前景很好，尤其是这些大学都培养了优秀的研究生。

DEMIS HASSABIS: Both really. I’m a born-and-bred Londoner, and I love London, but at the same time, I thought it was a competitive advantage because the UK and Europe have amazing universities in the field of AI like Cambridge and Oxford. But also, at the time there was no real ambitious research company in the UK, or really in Europe, so our hiring prospects were high, especially with all these universities outputting great postgraduate and graduate students.

2018 年，欧洲已经有多家公司，但我们是第一家在人工智能领域进行深入研究的公司。但从文化角度来看，我认为重要的是，我们需要更多利益相关者和文化参与到人工智能的开发中，不仅仅是美国的硅谷，还有欧洲和加拿大的情感等等。最终，这将具有全球意义，而关于如何使用人工智能、用在什么地方以及如何分配收益，拥有不同的声音非常重要。

In 2018 there are now a number of companies in Europe, but we were the first in AI who were doing deep research. But more culturally, I think it’s important that we have more stakeholders and cultures involved in making AI, not just Silicon Valley in the United States, but also European sensibilities and Canadian, and so on. Ultimately, this is going to be of global significance and having different voices about how to use it, what to use it for, and how to distribute the proceeds, is important.

马丁·福特：我相信你们也在其他欧洲城市开设了实验室？

MARTIN FORD: I believe you’re also opening up labs in other European cities?

德米斯·哈萨比斯：我们在巴黎开设了一个小型研究实验室，这是我们在欧洲大陆的第一个办事处。我们还在加拿大的阿尔伯塔省和蒙特利尔开设了两个实验室。最近，自加入谷歌以来，我们在加利福尼亚州山景城设立了一个应用团队办公室，就在我们合作的谷歌团队旁边。

DEMIS HASSABIS: We’ve opened a small research lab in Paris, which is our first continental European office. We’ve also opened two labs in Canada in Alberta and Montreal. More recently, since joining Google, we now have an applied team office in Mountain View, California who are right next to the Google teams that we work with.

马丁·福特：您与谷歌的其他人工智能团队的合作有多密切？

MARTIN FORD: How closely do you work with the other AI teams at Google?

DEMIS HASSABIS：谷歌规模庞大，有成千上万的人在研究机器学习和人工智能的各个方面，既有非常实用的视角，也有纯粹的研究视角。因此，有许多团队负责人都互相认识，而且有很多交叉合作，既有产品团队的，也有研究团队的。这往往是临时性的，所以这取决于个别研究人员或个别主题，但我们会在高层相互通报我们的整体研究方向。

DEMIS HASSABIS: Google’s a huge place, and there are thousands of people working on every aspect of machine learning and AI, from both a very applied perspective to a pure research point of view. As a result of that, there are a number of team leads who all know each other, and there’s a lot of cross-collaboration, both with product teams and research teams. It tends to be ad hoc, so it depends on individual researchers or individual topics, but we keep each other informed at a high level of our overall research directions.

在 DeepMind，我们与其他团队截然不同，因为我们非常专注于 AGI 这一宏伟目标。我们围绕长期路线图进行组织，这是我们基于神经科学的论文，它讨论了什么是智能以及实现智能需要什么。

At DeepMind, we’re quite different from other teams in that we’re pretty focused around this one moonshot goal of AGI. We’re organized around a long-term roadmap, which is our neuroscience-based thesis, which talks about what intelligence is and what’s required to get there.

马丁·福特：DeepMind 在 AlphaGo 上取得的成就有目共睹。甚至还有一部关于它的纪录片（ https://www.alphagomovie.com/），所以我想更多地关注你的最新创新 AlphaZero 以及你的未来计划。在我看来，你已经展示了一种非常接近信息完全双人游戏的通用解决方案；换句话说，在这种游戏中，所有已知的信息都可以在棋盘上或屏幕上的像素中找到。展望未来，你会完成这类游戏吗？你打算转向更复杂的隐藏信息游戏吗？等等？

MARTIN FORD: DeepMind’s accomplishments with AlphaGo are well documented. There’s even a documentary film about it (https://www.alphagomovie.com/), so I wanted to focus more on your latest innovation, AlphaZero, and on your plans for the future. It seems to me that you’ve demonstrated something very close to a general solution for information-complete two-player games; in other words, games where everything that can be known is available there on the board or in terms of pixels on the screen. Going forward, are you finished with that type of game? Are you planning to move on to more complex games with hidden information, and so forth?

DEMIS HASSABIS：我们即将发布一个经过进一步改进的新版 AlphaZero，正如您所说，您可以将其视为国际象棋、围棋、将棋等双人完美信息游戏的解决方案。当然，现实世界并不是由完美信息组成的，所以正如您所说，下一步是创建能够处理此类问题的系统。我们已经在研究这个问题，其中一个例子就是我们与 PC 战略游戏《星际争霸》的合作，这款游戏的动作空间非常复杂。它非常复杂，因为您需要构建单位，因此就您拥有的棋子而言，它并不像国际象棋那样是静态的。它也是实时的，游戏中有隐藏信息，例如“战争迷雾”会遮蔽屏幕上的信息，直到您探索该区域。

DEMIS HASSABIS: There’s a new version of AlphaZero that we’re going to publish soon that’s even more improved, and as you’ve said, you can think of that as a solution to two-player perfect-information games like chess, Go, shogi, and so on. Of course, the real world is not made up of perfect information, so as you’ve said, the next step is to create systems that can deal with that. We’re already working on that, and one example of this is our work with the PC strategy game, StarCraft, which has a very complicated action space. It’s very complex because you build units, so it’s not static in terms of what pieces you have, like in chess. It’s also real time, and the game has hidden information, for example, the “fog of war” that obscures onscreen information until you explore that area.

除此之外，游戏只是我们的训练领域。我们做这些工作不只是为了解决游戏问题；我们希望构建这些可以应用于现实世界问题的通用算法。

Beyond that, games are just our training domain. We’re not doing all this work just to solve games; we want to build these general algorithms that we can apply to real-world problems.

马丁·福特：到目前为止，您的重点主要是将深度学习与强化学习相结合。这基本上是通过实践进行学习，系统会反复尝试某件事，并且有一个奖励函数推动它走向成功。我听说您说过，您相信强化学习为通用智能提供了一条可行的途径，它可能足以实现这一目标。这是您未来的主要关注点吗？

MARTIN FORD: So far, your focus has primarily been on combining deep learning with reinforcement learning. That’s basically learning by practice, where the system repeatedly attempts something, and there’s a reward function that drives it toward success. I’ve heard you say that you believe that reinforcement learning offers a viable path to general intelligence, that it might be sufficient to get there. Is that your primary focus going forward?

DEMIS HASSABIS：是的，未来会这样。我认为这项技术非常强大，但你需要将它与其他技术结合起来才能扩大规模。强化学习已经存在很长时间了，但它只用于解决非常小的玩具问题，因为任何人都很难以任何方式扩大这种学习的规模。在我们的 Atari 工作中，我们将其与深度学习相结合，深度学习负责处理屏幕和你所处环境的模型。深度学习在扩展方面非常出色，因此将其与强化学习相结合使其能够扩展到我们现在在 AlphaGo 和 DQN 中解决的这些大问题——所有这些事情在 10 年前人们会告诉你是不可能的。

DEMIS HASSABIS: Going forward, yes, it is. I think that technique is extremely powerful, but you need to combine it with other things to scale it. Reinforcement learning has been around for a long time, but it was only used in very small toy problems because it was very difficult for anyone to scale up that learning in any way. In our Atari work, we combined that with deep learning, which did the processing of the screen, and the model of the environment you’re in. Deep learning is amazing at scaling, so combining that with reinforcement learning allowed it to scale to these large problems that we’ve now tackled in AlphaGo and DQN—all of these things that people would have told you was impossible 10 years ago.

我认为我们已经证明了第一部分。我们之所以如此有信心，并支持它，是因为我认为强化学习在未来几年将变得和深度学习一样重要。DeepMind 是少数认真对待这一问题的公司之一，因为从神经科学的角度来看，我们知道大脑使用一种强化学习作为其学习机制之一，它被称为时间差异学习，我们知道多巴胺系统实现了这一点。你的多巴胺神经元会跟踪大脑的预测错误，然后你根据这些奖励信号加强你的突触。大脑按照这些原则运作，大脑是我们唯一的通用智能例子，这就是我们在这里非常重视神经科学的原因。对我们来说，这一定是解决通用智能问题的可行方法。它可能不是唯一的，但从生物学启发的角度来看，只要你将其扩大到足够大，强化学习似乎就足够了。当然，这样做有很多技术挑战，其中许多尚未解决。

I think we proved that first part. The reason we were so confident about it and why we backed it when we did was because in my opinion reinforcement learning will become as big as deep learning in the next few years. DeepMind is one of the few companies that take that seriously because, from the neuroscience perspective, we know that the brain uses a form of reinforcement learning as one of its learning mechanisms, it’s called temporal difference learning, and we know the dopamine system implements that. Your dopamine neurons track the prediction errors your brain is making, and then you strengthen your synapses according to those reward signals. The brain works along these principles, and the brain is our only example of general intelligence, which is why we take neuroscience very seriously here. To us, that must be a viable solution to the problem of general intelligence. It may not be the only one, but from a biologically inspired standpoint, it seems reinforcement learning is sufficient once you scale it up enough. Of course, there are many technical challenges with doing that, and many of them are unsolved.

马丁·福特：尽管如此，当孩子学习语言或理解世界等东西时，它在很大程度上似乎并不像强化学习。它是无监督学习，因为没有人像我们对待 ImageNet 那样给孩子标记数据。然而，不知何故，一个年幼的孩子可以直接从环境中有机地学习。但它似乎更多地受到观察或与环境的随机互动的驱动，而不是通过带着特定目标的实践来学习。

MARTIN FORD: Still, when a child learns things like language or an understanding of the world, it doesn’t really seem like reinforcement learning for the most part. It’s unsupervised learning, as no one’s giving the child labeled data the way we would do with ImageNet. Yet somehow, a young child can learn organically directly from the environment. But it seems to be more driven by observation or random interaction with the environment rather than learning by practice with a specific goal in mind.

德米斯·哈萨比斯：儿童学习时会使用多种机制，而大脑不会只使用一种机制。儿童从父母、老师或同龄人那里获得监督学习，当他们只是在试验东西而没有目标时，他们会进行无监督学习。当他们做某事时，他们也会进行奖励学习和强化学习，并因此获得奖励。

DEMIS HASSABIS: A child learns with many mechanisms, it’s not like the brain only uses one. The child gets supervised learning from their parents, teachers, or their peers and they do unsupervised learning when they’re just experimenting with stuff, with no goal in mind. They also do reward learning and reinforcement learning when they do something, and they get a reward for it.

我们正在研究这三种方法，它们都是智能所必需的。无监督学习非常重要，我们正在研究它。这里的问题是，进化是否为我们设计了内在动机，最终成为奖励的代理，然后指导无监督学习？看看信息增益。有强有力的证据表明，获取信息本质上对你的大脑有益。

We work on all three of those, and they’re all going to be needed for intelligence. Unsupervised learning is hugely important, and we’re working on that. The question here is, are there intrinsic motivations that evolution has designed in us that end up being proxies for reward, which then guide the unsupervised learning? Just look at information gain. There is strong evidence showing that gaining information is intrinsically rewarding to your brain.

另一件事是寻求新奇。我们知道看到新奇的事物会释放大脑中的多巴胺，这意味着新奇本身就是一种奖励。从某种意义上说，我们大脑中化学存在的这些内在动机可能正在引导我们进行无结构游戏或无监督学习。如果大脑发现寻找信息和结构本身就是一种奖励，那么这对无监督学习来说就是一种非常有用的动机；不管怎样，你都会试图找到结构，而大脑似乎正在这样做。

Another thing would be novelty seeking. We know that seeing novel things releases dopamine in the brain, so that means novelty is intrinsically rewarding. In a sense, it could be that these intrinsic motivations that we have chemically in our brains are guiding what seems to us to be unstructured play or unsupervised learning. If the brain finds finding information and structure rewarding in itself, then that’s a hugely useful motivation for unsupervised learning; you’re just going to try and find structure, no matter what, and it seems like the brain is doing that.

根据你确定的奖励，其中一些可能是内在奖励，可以指导无监督学习。我发现在强化学习的框架中思考智能很有用。

Depending on what you determine as the reward, some of these things could be intrinsic rewards that could be guiding the unsupervised learning. I find that it is useful to think about intelligence in the framework of reinforcement learning.

马丁·福特：从您的谈话中可以明显看出，您对神经科学和计算机科学有着浓厚的兴趣。这种结合的方法对整个 DeepMind 来说也是如此吗？公司如何整合这两个领域的知识和人才？

MARTIN FORD: One thing that’s obvious from listening to you is that you combine a deep interest in both neuroscience and computer science. Is that combined approach true for DeepMind as a whole? How does the company integrate knowledge and talent from those two areas?

DEMIS HASSABIS：我对这两个领域的研究都处于中间位置，因为我在这两个领域都受过同等的训练。我想说 DeepMind 显然更偏向机器学习；然而，我们 DeepMind 最大的团队是由马特·波特维尼克 (Matt Botvinick) 领导的神经科学家组成的，他是一位了不起的神经科学家，也是普林斯顿大学的教授。我们非常重视这件事。

DEMIS HASSABIS: I’m definitely right in the middle for both those fields, as I’m equally trained in both. I would say DeepMind is clearly more skewed towards machine learning; however, our biggest single group here at DeepMind is made up of neuroscientists led by Matt Botvinick, an amazing neuroscientist and professor from Princeton. We take it very seriously.

神经科学的问题在于它本身就是一个巨大的领域，比机器学习大得多。如果你是一名机器学习人员，想要快速找出神经科学的哪些部分对你有用，那么你就会陷入困境。没有书会告诉你这一点，只有大量的研究工作，你必须自己弄清楚如何解析这些信息，并从人工智能的角度找到有用的信息。大多数神经科学研究都是为医学研究、心理学或神经科学本身而进行的。神经科学家在设计这些实验时并没有想到它们会对人工智能有用。99% 的文献对你作为一名人工智能研究人员来说毫无用处，所以你必须非常善于训练自己，以便找到正确的影响因素以及每种影响因素的正确影响程度。

The problem with neuroscience is that it’s a massive field in itself, way bigger than machine learning. If you as a machine-learning person wanted to quickly find out which parts of neuroscience would be useful to you, then you’d be stuck. There’s no book that’s going to tell you that, there’s just a mass of research work, and you’ll have to figure out for yourself how to parse that information and find the nuggets that could be useful from an AI perspective. Most of that neuroscience research is being undertaken for medical research, psychology, or for neuroscience itself. Neuroscientists aren’t designing those experiments thinking they would be useful for AI. 99% of that literature is not useful to you as an AI researcher and so you have to get really good at training yourself to navigate and pick out what are the right influences and what is the right level of influence for each of those.

很多人都在谈论神经科学如何启发人工智能研究，但我认为他们中很多人并没有真正想到具体如何做到这一点。让我们来探讨两个极端。一个是你可以尝试对大脑进行逆向工程，这也是很多人在研究人工智能时尝试做的事情，我的意思是从皮层层面对大脑进行逆向工程，一个典型的例子就是蓝脑计划。

Quite a lot of people talk about neuroscience inspiring AI work, but I don’t think a lot of them really have concrete ideas on how to do that. Let’s explore two extremes. One is you could try and reverse-engineer the brain, which is what quite a lot of people are attempting to do in their approach to AI, and I mean literally reverse-engineer the brain on a cortical level, a prime example being the Blue Brain Project.

马丁·福特：这是亨利·马克拉姆执导的，对吗？

MARTIN FORD: That’s being directed by Henry Markram, right?

DEMIS HASSABIS：是的，他实际上是在尝试对皮质柱进行逆向工程。这可能是一门有趣的神经科学，但在我看来，这不是构建人工智能的最有效途径，因为它太低级了。我们在 DeepMind 感兴趣的是对大脑和大脑实施的算法、它所拥有的能力、它所拥有的功能以及它所使用的表示的系统级理解。

DEMIS HASSABIS: Right, and he’s literally trying to reverse-engineer cortical columns. It may be interesting neuroscience but, in my view, that is not the most efficient path towards building AI because it’s too low-level. What we’re interested in at DeepMind is a systems-level understanding of the brain and the algorithms the brain implements, the capabilities it has, the functions it has, and the representations it uses.

DeepMind 并没有研究湿件的具体细节，也没有研究生物学如何将其实例化，我们可以将所有这些都抽象出来。这是有道理的，因为你为什么会想到一个计算机系统必须模仿一个计算机系统，因为这两个系统有完全不同的优点和缺点。在硅片上，你没有理由想要复制海马体的精确排列细节。另一方面，我对海马体的计算和功能非常感兴趣，比如情景记忆、空间导航和它使用的网格单元。这些都是来自神经科学的系统级影响，展示了我们对大脑使用的功能、表示和算法的兴趣，而不是实现的精确细节。

DeepMind is not looking at the exact specifics of the wetware or how the biology actually instantiates it, we can abstract all of that away. That makes sense, because why would you imagine an in-silico system would have to mimic an in-carbo system because there are completely different strengths and weaknesses about those two systems. In silicon, there’s no reason why you would want to copy the exact permutation details of, say a hippocampus. On the other hand, I am very interested in the computations and the functions that the hippocampus has, like episodic memory, navigating in space, and the grid cells it uses. These are all systems-level influences from neuroscience and showcase our interest in the functions, representations and the algorithms that the brain uses, not the exact details of implementation.

马丁·福特：你经常听到这样的比喻：飞机不会扇动翅膀。飞机可以飞行，但不能精确模仿鸟类的动作。

MARTIN FORD: You often hear the analogy that airplanes don’t flap their wings. Airplanes achieve flight, but don’t precisely mimic what birds do.

德米斯·哈萨比斯：这是一个很好的例子。在 DeepMind，我们试图通过观察鸟类来理解空气动力学，然后抽象出空气动力学的原理并建造一架固定翼飞机。

DEMIS HASSABIS: That’s a great example. At DeepMind, we’re trying to understand aerodynamics by looking at birds, and then abstracting the principles of aerodynamics and building a fixed-wing plane.

当然，制造飞机的人是受到鸟类的启发。莱特兄弟知道比空气重的飞行是可能的，因为他们见过鸟类。在机翼发明之前，他们曾尝试使用可变形的机翼，但没有成功，但它们更像滑翔的鸟类。你要做的就是观察自然，然后试着抽象出那些对你追求的现象（飞行，在我们的例子中是智能）不重要的东西。但这并不意味着这对你的搜索过程没有帮助。

Of course, people who built planes were inspired by birds. The Wright Brothers knew that heavier-than-air flight was possible because they’d seen birds. Before the airfoil was invented, they tried without success to use deformable wings, but they were more like birds gliding. What you’ve got to do is look at nature, and then try and abstract away the things that are not important for the phenomenon you’re after in that case, flying and in our case, intelligence. But that doesn’t mean that that didn’t help your search process.

我的观点是，你还不知道结果会是什么样的。如果你试图构建人工智能之类的人工智能，但它不能立即发挥作用，你怎么知道你找对了地方？你的 20 人团队是在浪费时间吗？还是你应该再努力一点，也许明年你就能破解它？正因为如此，以神经科学为指导，我可以对这类事情下更大、更有力的赌注。

My point is that you don’t know yet what the outcome looks like. If you’re trying to build something artificial like intelligence and it doesn’t work straight away, how do you know that you’re looking in the right place? Is your 20-person team wasting their time, or should you push a bit harder, and maybe you’ll crack it next year? Because of that, having neuroscience as a guide can allow me to make much bigger, much stronger bets on things like that.

强化学习就是一个很好的例子。我知道强化学习必须是可扩展的，因为大脑确实可以对其进行扩展。如果你不知道大脑实现了强化学习，而且它无法扩展，你怎么知道在实践层面上你是否应该再花两年时间研究它？缩小团队或公司探索的搜索空间非常重要，我认为这是忽视神经科学的人经常忽略的一个基本点。

A great example of this is reinforcement learning. I know reinforcement learning has to be scalable because the brain does scale it. If you didn’t know that the brain implemented reinforcement learning and it wasn’t scaling, how would you know on a practical level if you should spend another two years on this? It’s very important to narrow down the search space that you’re exploring as a team or a company, and I think that’s a meta-point that is often missed by people that ignore neuroscience.

马丁·福特：我认为你已经指出，人工智能领域的研究也可以为神经科学领域的研究提供参考。DeepMind 刚刚公布了用于导航的网格细胞的研究结果，听起来你已经让它们自然地出现在神经网络中。换句话说，相同的基本结构自然出现在生物大脑和人工神经网络中，这似乎相当了不起。

MARTIN FORD: I think you’ve made the point that the work in AI could also inform research being done in neuroscience. DeepMind just came out with a result on grid cells used in navigation, and it sounds like you’ve got them to emerge organically in a neural network. In other words, the same basic structure naturally arises in both the biological brain and in artificial neural networks, which seems pretty remarkable.

DEMIS HASSABIS：我对此感到非常兴奋，因为这是我们去年取得的最大突破之一。发现网格细胞并因此获得诺贝尔奖的 Edvard Moser 和 May-Britt Moser 都写信给我们，对这一发现感到非常兴奋，因为这意味着这些网格细胞可能不仅仅是大脑布线的功能，实际上可能是从计算意义上表示空间的最佳方式。这对神经科学家来说是一个重大的发现，因为他们现在推测，也许大脑不一定天生就具有创建网格细胞的能力。也许如果你有这种神经元结构，并将它们暴露在空间中，这就是任何系统都能想出的最有效的编码。

DEMIS HASSABIS: I’m very excited about that because it’s one of our biggest breakthroughs in the last year. Edvard Moser and May-Britt Moser, who discovered grid cells and won the Nobel Prize for their work both wrote to us very excited about this finding because it means that, possibly, these grid cells are not just a function of the wiring of the brain, but actually may be the most optimal way of representing space from a computational sense. That’s a huge and important finding for the neuroscientists because what they’re speculating now is that maybe the brain isn’t necessarily hardwired to create grid cells. Perhaps if you have that structure of neurons and you just expose them to space, that is the most efficient coding any system would come up with.

我们最近还根据对我们的人工智能算法及其作用的观察，创建了有关前额叶皮层如何运作的全新理论，然后让我们的神经科学家将其转化为大脑可能如何运作。

We’ve also recently created a whole new theory around how the prefrontal cortex might work, based on looking at our AI algorithms and what they were doing, and then having our neuroscientists translate that into how the brain might work.

我认为，这是看到更多人工智能思想和算法的例子的开始，这些思想和算法激励我们以不同的方式看待大脑中的事物或在大脑中寻找新事物，或将其作为分析工具来试验我们对大脑如何运作的想法。

I think that this is the beginning of seeing many more examples of AI ideas and algorithms inspiring us to look at things in a different way in the brain or looking for new things in the brain, or as an analysis tool to experiment with our ideas about how we think the brain might work.

作为一名神经科学家，我认为我们正在构建受神经科学启发的人工智能，这是解决一些关于大脑的复杂问题的最佳方式之一。如果我们建立一个基于神经科学的人工智能系统，我们就可以将其与人类大脑进行比较，并可能开始收集一些关于其独特特征的信息。我们可以开始揭示一些关于心灵的深奥奥秘，比如意识、创造力和梦想的本质。我认为将大脑与算法结构进行比较可以理解这一点。

As a neuroscientist, I think that the journey we’re on of building neuroscience-inspired AI is one of the best ways to address some of the complex questions we have about the brain. If we build an AI system that’s based on neuroscience, we can then compare it to the human brain and maybe start gleaning some information about its unique characteristics. We could start shedding light on some of the profound mysteries of the mind like the nature of consciousness, creativity, and dreaming. I think that comparing the brain to an algorithmic construct could be a way to understand that.

马丁·福特：听起来你认为可能存在一些可发现的与基质无关的智能一般原则。回到飞行类比，你可以称之为“智能的空气动力学”。

MARTIN FORD: It sounds like you think there could be some discoverable general principles of intelligence that are substrate-independent. To return to the flight analogy, you might call it “the aerodynamics of intelligence.”

德米斯·哈萨比斯：没错，如果你提取出这个一般原则，那么它对于理解人类大脑的特定情况一定很有用。

DEMIS HASSABIS: That’s right, and if you extract that general principle, then it must be useful for understanding the particular instance of the human brain.

马丁·福特：您能谈谈您想象中未来 10 年内会发生的一些实际应用吗？在不久的将来，您的突破将如何应用于现实世界？

MARTIN FORD: Can you talk about some of the practical applications that you imagine happening within the next 10 years? How are your breakthroughs going to be applied in the real world in the relatively near future?

DEMIS HASSABIS：我们已经在实践中看到了很多成果。如今，世界各地的人们都通过机器翻译、图像分析和计算机视觉与人工智能互动。

DEMIS HASSABIS: We’re already seeing lots of things in practice. All over the world people are interacting with AI today through machine translation, image analysis, and computer vision.

DeepMind 已经开始研究很多事情，比如优化谷歌数据中心的能源使用。我们研究了 WaveNet，这是一种非常人性化的文本转语音系统，现在所有 Android 手机的 Google Assistant 都内置了该系统。我们在推荐系统、Google Play 甚至后台元素（比如节省 Android 手机的电池寿命）中使用人工智能。这些都是每个人每天都在使用的东西。我们发现，由于它们是通用算法，因此它们无处不在，所以我认为这只是个开始。

DeepMind has started working on quite a few things, like optimizing the energy being used in Google’s data centers. We’ve worked on WaveNet, the very human-like text-to-speech system that’s now in the Google Assistant in all Android-powered phones. We use AI in recommendation systems, in Google Play, and even on behind-the-scenes elements like saving battery life on your Android phone. Things that everyone uses every single day. We’re finding that because they’re general algorithms, they’re coming up all over the place, so I think that’s just the beginning.

我希望接下来我们在医疗保健领域的合作能够顺利实现。例如，我们与英国著名的眼科医院 Moorfields 合作，希望通过视网膜扫描诊断黄斑变性。我们在《自然医学》杂志上发表了我们联合研究合作第一阶段的结果，结果表明，我们的人工智能系统能够以前所未有的准确度快速解读常规临床实践中的眼部扫描结果。它还能像世界领先的专家医生一样，准确地推荐患者如何转诊治疗 50 多种威胁视力的眼部疾病。

What I’m hoping will come through next are the collaborations we have in healthcare. An example of this is our work with the famous UK eye hospital, Moorfields, where we’re looking at diagnosing macular degeneration from your retina scans. We published the results from the first phase of our joint research partnership in Nature Medicine, and they show that our AI system can quickly interpret eye scans from routine clinical practice with unprecedented accuracy. It can also correctly recommend how patients should be referred for treatment for over 50 sight-threatening eye diseases as accurately as world-leading expert doctors.

还有其他团队也在为皮肤癌等疾病开展类似的研究。在未来五年内，我认为医疗保健将成为从我们在该领域所做的工作中获益最大的领域之一。

There are other teams doing similar work for diseases like skin cancer. Over the next five years, I think healthcare will be one of the biggest areas to see a benefit from the work we’re all doing in the field.

我个人真正感到兴奋的是，我认为我们正处于一个转折点，那就是使用人工智能来解决科学问题。我们正在研究蛋白质折叠等问题，但你可以想象它在材料设计、药物发现和化学方面的应用。人们正在使用人工智能分析大型强子对撞机的数据，以寻找系外行星。有很多非常酷的海量数据领域，我们作为人类专家很难识别其中的结构，我认为这种人工智能将得到越来越多的应用。我希望在未来 10 年内，这将加速一些真正基础领域的科学突破。

What I’m really personally excited about, and this is something I think we’re on the cusp of, is using AI to actually help with scientific problems. We’re working on things like protein folding, but you can imagine its use in material design, drug discovery and chemistry. People are using AI to analyze data from the Large Hadron Collider to searching for exoplanets. There’s a lot of really cool areas of masses of data that we as human experts find hard to identify the structure in that I think this kind of AI is going to become increasingly used for. I’m hoping that over the next 10 years this will result in an advancement in the speed of scientific breakthroughs in some really fundamental areas.

马丁·福特：AGI 的发展之路是什么样的？您认为在实现人类级别的 AI 之前，我们必须克服的主要障碍是什么？

MARTIN FORD: What does the path to AGI look like? What would you say are the main hurdles that will have to be surmounted before we have human-level AI?

DEMIS HASSABIS：从 DeepMind 成立之初，我们就确定了一些重要的里程碑，比如学习抽象的概念知识，然后将其用于迁移学习。迁移学习就是将知识从一个领域有效地迁移到一个你从未见过的新领域，这是人类的强项。如果你给我一个新任务，我不会表现得很糟糕，因为我会从类似的东西或结构性的东西中汲取一些知识，我可以立即开始处理它。这是计算机系统非常不擅长的事情，因为它们需要大量数据，而且效率很低。我们需要改进这一点。

DEMIS HASSABIS: From the beginning of DeepMind we identified some big milestones, such as the learning of abstract, conceptual knowledge, and then using that for transfer learning. Transfer learning is where you usefully transfer your knowledge from one domain to a new domain that you’ve never seen before, it’s something humans are amazing at. If you give me a new task, I won’t be terrible at it out of the box because I’ll bring some knowledge from similar things or structural things, and I can start dealing with it straight away. That’s something that computer systems are pretty terrible at because they require lots of data and they’re very inefficient. We need to improve that.

另一个里程碑是我们需要提高语言理解能力，另一个是使用我们的新技术复制旧人工智能系统能够做的事情，比如符号操作。我们距离所有这些目标还有很长的路要走，但如果这些目标实现了，那将是真正的里程碑。如果你看看我们在 2010 年，也就是八年前的情况，我们已经取得了一些对我们来说具有里程碑意义的大成就，比如 AlphaGo，但未来还会有更多。所以对我来说，这些是最重要的，概念和迁移学习。

Another milestone is that we need to get better at language understanding, and another is replicating things that old AI systems were able to do, like symbolic manipulation, but using our new techniques. We’re a long way from all of those, but they would be really big milestones if they were to happen. If you look at where we were in 2010, just eight years ago, we’ve already achieved some big things that were milestones to us, like AlphaGo, but there are more to come. So those would be the big ones for me, concepts and transfer learning.

马丁·福特：当我们实现 AGI 时，您是否想象过智能与意识相结合？它是会自动出现的东西，还是意识是完全独立的东西？

MARTIN FORD: When we do achieve AGI, do you imagine intelligence being coupled with consciousness? Is it something that would automatically emerge, or is consciousness a completely separate thing?

德米斯·哈萨比斯：这是这次旅程将要解决的一个有趣问题。我目前还不知道答案，但这是我们和其他人在这个领域所做的工作中最令人兴奋的事情之一。

DEMIS HASSABIS: That’s one of the interesting questions that this journey will address. I don’t know the answer to it at the moment, but that’s one of the very exciting things about the work that both we and others are doing in this field.

我目前的直觉是意识和智力是可以双重分离的。你可以拥有智力而没有意识，也可以拥有意识而没有人类水平的智力。我非常肯定聪明的动物有一定程度的意识和自我意识，但它们显然没有那么聪明，至少与人类相比是这样，我可以想象建造出从某种程度上来说非常聪明但对我们来说根本不会有任何意识的机器。

My hunch currently would be that consciousness and intelligence are double-dissociable. You can have intelligence without consciousness, and you can have consciousness without human-level intelligence. I’m pretty sure smart animals have some level of consciousness and self-awareness, but they’re obviously not that intelligent at least compared to humans, and I can imagine building machines that are phenomenally intelligent by some measures but would not feel conscious to us in any way at all.

马丁·福特：就像一个聪明的僵尸，没有内在体验。

MARTIN FORD: Like an intelligent zombie, something that has no inner experience.

DEMIS HASSABIS：某种东西不会像我们对其他人类的感觉那样具有知觉。这是一个哲学问题，因为问题在于，正如我们在图灵测试中看到的那样，我们如何知道它的行为是否与我们相同？奥卡姆剃刀的解释是，如果你表现出和我相同的行为，你和我是由相同的材料制成的，我知道我的感受，那么我可以假设你和我有相同的感受。你为什么不呢？

DEMIS HASSABIS: Something that wouldn’t feel sentient in the way we feel about other humans. Now that’s a philosophical question, because the problem is, as we see with the Turing test, how would we know if it was behaving in the same way as we were? The Occam’s razor explanation is to say that if you’re exhibiting the same behavior as I exhibit, and you’re made from the same stuff as I’m made from, and I know what I feel, then I can assume you’re feeling the same thing as me. Why would you not?

机器的有趣之处在于，如果我们设计它们，它们可以表现出与人类相同的行为，但它们处于不同的基底上。如果你不在同一基底上，那么奥卡姆剃刀的理念就不那么站得住脚了。也许它们在某种意义上是有意识的，但我们感觉不一样，因为我们没有可以依赖的额外假设。如果你分析一下为什么我们认为我们每个人都有意识，我认为这是一个非常重要的假设，如果你和我在同一基底上操作，为什么感觉会与你的基底不同？

What’s interesting with a machine is that they could exhibit the same behavior as a human, if we designed them like that, but they’re on a different substrate. If you’re not on the same substrate then that Occam’s razor idea doesn’t hold as strongly. It may be that they are conscious in some sense, but we don’t feel it in the same way because we don’t have that additional assumption to rely on. If you break down why we think each of us is conscious, I think that’s a very important assumption, if you’re operating on the same substrate as me, why would it feel different to your substrate?

马丁·福特：你相信机器意识是可能的吗？有些人认为意识从根本上来说是一种生物现象。

MARTIN FORD: Do you believe machine consciousness is possible? There are some people that argue consciousness is fundamentally a biological phenomenon.

DEMIS HASSABIS：我对此持开放态度，因为我认为我们对此一无所知。生物系统可能存在一些非常特殊的东西。有些人，比如罗杰·彭罗斯爵士，认为这与量子意识有关，在这种情况下，传统计算机不会拥有这种意识，但这仍是一个悬而未决的问题。这就是为什么我认为我们所走的道路将对此有所启发，因为我认为我们实际上不知道这是否是一个极限。无论如何，这都将令人着迷，因为如果事实证明你根本无法在机器上建立意识，那将非常令人惊奇。这将告诉我们很多关于意识是什么以及它存在于何处的信息。

DEMIS HASSABIS: I am actually open-minded about that, in the sense that I don’t think we know. It could well turn out that there’s something very special about biological systems. There are people like Sir Roger Penrose that think it’s to do with quantum consciousness, in which case a classical computer wouldn’t have it, but it’s an open question. That’s why I think the path we’re on will shed some light on it because I actually think we don’t know whether that’s a limit or not. Either way, it will be fascinating because it would be pretty amazing if it turned out that you couldn’t build consciousness at all on a machine. That would tell us a lot about what consciousness is and where it resides.

马丁·福特：AGI 的风险和弊端又如何呢？埃隆·马斯克曾谈到“培养恶魔”和生存威胁。还有尼克·博斯特罗姆，我知道他是 DeepMind 的顾问委员会成员，并就这个想法写了很多文章。您如何看待这些担忧？我们应该担心吗？

MARTIN FORD: What about the risks and the downsides associated with AGI? Elon Musk has talked about “raising the demon” and an existential threat. There’s also Nick Bostrom, who I know is on DeepMind’s advisory board and has written a lot on this idea. What do you think about these fears? Should we be worried?

德米斯·哈萨比斯：我和他们谈过很多这些事情。一如既往，这些言论听起来有些极端，但当你和这些人当面交谈时，你会发现情况要微妙得多。

DEMIS HASSABIS: I’ve talked to them a lot about these things. As always, the soundbites seem extreme but it’s a lot more nuanced when you talk to any of these people in person.

我对此的看法是，我处于中间位置。我之所以研究人工智能是因为我认为它将是人类有史以来最有益的东西。我认为它将以各种方式释放我们在科学和医学领域的潜力。与任何强大的技术一样，人工智能可能特别强大，因为它非常通用，技术本身是中性的。这取决于我们人类如何决定设计和部署它，我们决定使用它做什么，以及我们决定如何分配收益。

My view on it is that I’m in the middle. The reason I work on AI is because I think it’s going to be the most beneficial thing to humanity ever. I think it’s going to unlock our potential within science and medicine in all sorts of ways. As with any powerful technology, and AI could be especially powerful because it’s so general, the technology itself is neutral. It depends on how we as humans decide to design and deploy it, what we decide to use it for, and how we decide to distribute the gains.

这其中有很多复杂因素，但这些更像是地缘政治问题，需要我们作为一个社会来解决。尼克·博斯特罗姆担心的很多问题都是我们必须解决的技术问题，例如控制问题和价值观一致问题。我的观点是，在这些问题上我们确实需要进行更多的研究，因为我们现在才刚刚开始研究可以做任何有趣的事情的系统。

There are a lot of complications there, but those are more like geopolitical issues that we need to solve as a society. A lot of what Nick Bostrom worries about are the technical questions we have to get right, such as the control problem and the value alignment problem. My view is that on those issues we do need a lot more research because we’ve only just got to the point now where there are systems that can even do anything interesting at all.

我们仍处于非常早期的阶段。五年前，你可能还在谈论哲学，因为没有人提出任何有趣的东西。我们现在有了 AlphaGo 和其他一些仍然处于早期阶段的有趣技术，但我们现在应该开始对这些东西进行逆向工程，并通过构建可视化和分析工具对它们进行实验。我们已经让团队这样做，以更好地了解这些黑箱系统在做什么以及我们如何解释它们的行为。

We’re still at a very nascent stage. Five years ago, you might as well have been talking about philosophy because no one had anything that was interesting. We’ve now got AlphaGo and a few other interesting technologies that are still very nascent, but we’re now at the point where we should start reverse-engineering those things and experimenting on them by building visualization and analysis tools. We’ve got teams doing this to better understand what these black-box systems are doing and how we interpret their behavior.

马丁·福特：您是否有信心我们能够管理先进人工智能带来的风险？

MARTIN FORD: Are you confident that we’ll be able to manage the risks that come along with advanced AI?

DEMIS HASSABIS：是的，我非常有信心，原因是我们正处于一个转折点，我们刚刚让这些东西运转起来，还没有投入太多精力去对它们进行逆向工程和理解，而现在这种情况正在发生。在未来十年，大多数这些系统将不再是我们现在所指的黑匣子。我们将很好地掌握这些系统的运行情况，这将使我们更好地了解如何控制系统以及它们的数学极限是什么，然后这可以带来最佳实践和协议。

DEMIS HASSABIS: Yes, I’m very confident, and the reason is that we’re at the inflection point where we’ve just got these things working, and not that much effort has yet gone into reverse engineering them and understanding them, and that’s happening now. Over the next decade, most of these systems won’t be black-box in the sense that we mean now. We’ll have a good handle on what’s going on with these systems, and that will lead to a better understanding of how to control the systems and what their limits are mathematically, and then that could lead into best practices and protocols.

我非常有信心，这条道路将解决尼克·博斯特罗姆等人担心的很多技术问题，比如目标设定不正确的附带后果。为了在这方面取得进展，我一直认为，最好的科学是理论和实践——实证工作——齐头并进，而对于这个主题和领域来说，实证工作实验就是工程。

I’m pretty confident that path will address a lot of the technical issues that people like Nick Bostrom are worried about, like the collateral consequences of goals not being set correctly. To make advances in that, my view has always been that the best science occurs when theory and practice—empirical work—go hand in hand, and for this subject and field, empirical work experiments are engineering.

一旦我们真正更好地了解这些系统，许多不在该技术领域工作的人的担忧将不复存在。这并不是说我认为没有什么可担心的，因为我认为我们应该担心这些事情。还有很多短期问题需要解决——比如，在将这些系统部署到产品中时，我们如何测试它们？一些长期问题非常棘手，我们希望利用现在的时间思考这些问题，在我们需要答案之前就考虑它们。

A lot of the fears that some of the people not working at the coalface of this technology have won’t hold once we actually have a much better understanding of these systems. That’s not to say that I think that there’s nothing to worry about, because I think we should worry about these things. There are plenty of near-term questions to resolve as well—like how do we test these systems as we deploy them in products? Some of the long-term problems are so hard that we want to be thinking about them in the time we have right now, well ahead of when we’re going to need the answers.

我们还需要能够为必须进行的研究提供信息，以便找到 Nick Bostrom 等人提出的一些问题的解决方案。我们正在积极思考这些问题，并认真对待它们，但我坚信，如果全世界共同投入足够的智慧，人类的智慧一定能够克服这些问题。

We also need to be able to inform the research that has to be done to come up with the solutions to some of those questions that are posed by people like Nick Bostrom. We are actively thinking about these problems and we’re taking them seriously, but I’m a big believer in human ingenuity to overcome those problems if you put enough brainpower on it collectively around the world.

马丁·福特：在实现 AGI 之前，还会出现哪些风险呢？例如，自动武器。我知道您一直直言不讳地谈论人工智能在军事领域的应用。

MARTIN FORD: What about the risks that will arise long before AGI is achieved? For example, autonomous weapons. I know you’ve been very outspoken about AI being used in military applications.

DEMIS HASSABIS：这些都是非常重要的问题。在 DeepMind，我们以这样的前提为出发点：人工智能应用应处于人类有意义的控制之下，并用于有益于社会的目的。这意味着禁止开发和部署全自动武器，因为这需要人类具有一定程度的判断力和控制力，以确保武器的使用方式是必要且适当的。我们通过多种方式表达了这一观点，包括签署一封公开信和支持未来生命研究所就此问题做出的承诺。

DEMIS HASSABIS: These are very important questions. At DeepMind, we start from the premise that AI applications should remain under meaningful human control, and be used for socially beneficial purposes. This means banning the development and deployment of fully autonomous weapons, since it requires a meaningful level of human judgment and control to ensure that weapons are used in ways that are necessary and proportionate. We’ve expressed this view in a number of ways, including signing an open letter and supporting the Future of Life Institute’s pledge on the subject.

马丁·福特：值得指出的是，尽管化学武器已被禁止，但仍然有人使用过。所有这些都需要全球协调，而各国之间的竞争似乎可能会将事情推向另一个方向。例如，人们认为美国与中国在人工智能领域展开了竞争。他们的政府体系确实更加专制。我们是否应该担心他们会在人工智能领域取得优势？

MARTIN FORD: It’s worth pointing out even though chemical weapons are in fact banned, they have still been used. All of this requires global coordination and it seems that rivalries between countries could push things in the other direction. For example, there is a perceived AI race with China. They do have a much more authoritarian system of government. Should we worry that they will gain an advantage in AI?

DEMIS HASSABIS：从这个意义上来说，我不认为这是一场竞赛，因为我们认识所有的研究人员，而且有很多合作。我们公开发表论文，我知道腾讯已经创建了一个 AlphaGo 克隆版，所以我认识那里的许多研究人员。我确实认为，如果将来要进行协调，甚至制定监管和最佳实践，那么重要的是它是国际性的，全世界都采用它。如果一些国家不采用这些原则，它就行不通。然而，这并不是人工智能独有的问题。我们已经在努力解决许多其他问题，这些问题是全球协调和组织的问题——最明显的一个就是气候变化。

DEMIS HASSABIS: I don’t think it’s a race in that sense because we know all the researchers and there’s a lot of collaboration. We publish papers openly and I know that for example Tencent has created an AlphaGo clone, so I know many of the researchers there. I do think that if there’s going to be coordination and perhaps even regulation and best practices down the road, it’s important that it’s international and the whole world adopts that. It doesn’t work if some countries don’t adopt those principles. However, that’s not an issue that’s unique to AI. There are many other problems that we’re already grappling with that are a question of global coordination and organization—the obvious one being climate change.

马丁·福特：这一切对经济有什么影响？就业市场是否会受到严重影响，失业率和不平等现象是否会上升？

MARTIN FORD: What about the economic impact of all of this? Is there going to be a big disruption of the job market and perhaps rising unemployment and inequality?

德米斯·哈萨比斯：我认为到目前为止，人工智能带来的颠覆非常小，它只是技术颠覆的一部分。不过，人工智能将带来巨大的变革。有些人认为，人工智能的规模将与工业革命或电力相当，而另一些人则认为，人工智能将超越工业革命或电力，我认为这还有待观察。也许这意味着我们处在一个富足的世界，到处都有巨大的生产力增长？没有人知道确切答案。关键是要确保每个人都能分享这些好处。

DEMIS HASSABIS: I think there’s been very minimal disruption so far from AI, it’s just been part of the technology disruption in general. AI is going to be hugely transformative, though. Some people believe that it’s going to be on the scale of the Industrial Revolution or electricity, while other people believe it’s going to be a class of its own above that, and that’s something I think that remains to be seen. Maybe it will mean we’re in a world of abundance, where there are huge productivity gains everywhere? Nobody knows for sure. The key thing is to make sure those benefits are shared with everyone.

我认为这是关键，无论是全民基本收入还是其他形式。很多经济学家都在争论这些事情，我们需要非常仔细地思考社会中的每个人将如何从这些可能巨大的生产力增长中受益，这些增长肯定是存在的，否则它就不会造成如此大的破坏。

I think that’s the key thing, whether that’s universal basic income, or it’s done in some other form. There are lots of economists debating these things, and we need to think very carefully about how everyone in society will benefit from those presumably huge productivity gains, which must be coming in, otherwise it wouldn’t be so disruptive.

马丁·福特：是的，这基本上就是我的观点，从根本上说，这是一个分配问题，我们很大一部分人口面临着落后的危险。但要提出一种新的模式来创造一个为所有人服务的经济，这是一个艰巨的政治挑战。

MARTIN FORD: Yes, that’s basically the argument that I’ve been making, that it’s fundamentally a distributional problem and that a large part of our population is in danger of being left behind. But it is a staggering political challenge to come up with a new paradigm that will create an economy that works for everyone.

德米斯·哈萨比斯：对。

DEMIS HASSABIS: Right.

每当我遇到经济学家时，我都认为他们应该非常努力地解决这个问题，但这很难，因为他们无法真正想象它如何能如此富有成效，因为人们已经谈论了 100 年来大规模的生产力增长。

Whenever I meet an economist, I think they should be working quite hard on this problem, but it’s difficult to because they can’t really envisage how it could be so productive because people have been talking about massive productivity gains for 100 years.

我父亲在大学学习经济学，他说在 20 世纪 60 年代末，很多人都在认真讨论这个问题：“当我们拥有如此丰富的资源，而不需要工作时，每个人会做什么？”当然，这种情况从未发生过，无论是在 20 世纪 80 年代还是在那之后，我们比以往任何时候都更加努力地工作。我认为很多人不确定这种情况是否会永远持续下去，但如果我们最终拥有大量额外的资源和生产力，那么我们必须广泛而公平地分配它，我认为如果我们这样做，那么我认为这就没有什么问题。

My dad studied economics at university, and he was saying that in the late 1960s a lot of people were seriously talking about that: “What is everyone going to do in the 1980s when we have so much abundance, and we don’t have to work?” That, of course, never happened, in the 1980s or since then, and we’re working harder than ever. I think a lot of people are not sure if it’s ever going to be like that, but if it does end up that we have a lot of extra resources and productivity, then we’ve got to distribute it widely and equitably, and I think if we do that, then I don’t see a problem with it.

马丁·福特：您是个乐观主义者吗？我猜您认为人工智能具有变革性，而且可以说它是人类历史上最伟大的事情之一。当然，前提是我们能够明智地管理它。

MARTIN FORD: Is it safe to say that you’re an optimist? I’d guess that you see AI as transformative and that it’s arguably going to be one of the best things that’s ever happened to humanity. Assuming, of course, that we manage it wisely?

DEMIS HASSABIS：当然，这也是我一生都在为此努力的原因。我们在讨论的第一部分中提到的我所做的一切都是为了实现这一目标。如果没有人工智能，我会对世界的发展方向感到相当悲观。事实上，我认为世界上有很多问题需要更好的解决方案，比如气候变化、阿尔茨海默氏症研究或水净化。我可以给你列出一份会随着时间的推移而恶化的事情的清单。令人担忧的是，我不知道我们将如何获得全球协调以及多余的资源或活动来解决这些问题。但最终，我对世界还是持乐观态度的，因为像人工智能这样的变革性技术即将到来。

DEMIS HASSABIS: Definitely, and that’s why I’ve worked towards it my whole life. All of the things I’ve been doing that we covered in the first part of our discussion have been building towards achieving that. I would be quite pessimistic about the way the world’s going, if AI was not going to come along. I actually think there’s a lot of problems in the world that require better solutions, like climate change, Alzheimer’s research or water purification. I can give you a list of things that are going to get worse over time. What is a worry is that I don’t see how we’re going to get the global coordination and the excess resources or activity to solve them. But ultimately, I’m actually optimistic about the world because a transformative technology like AI is coming.

德米斯·哈萨比斯 曾是一名儿童国际象棋神童，他提前两年完成了高中考试，然后在 17 岁时编写了销量数百万的模拟游戏《主题公园》。在剑桥大学以计算机科学双一等学位毕业后，他创立了先锋视频游戏公司 Elixir Studios，为 Vivendi Universal 等全球发行商制作屡获殊荣的游戏。在领导成功的技术初创公司十年后，德米斯重返学术界，在伦敦大学学院攻读认知神经科学博士学位，随后在麻省理工学院和哈佛大学从事博士后研究。他对想象和规划背后的神经机制的研究被《科学》杂志列为 2007 年十大科学突破之一。

DEMIS HASSABIS is a former child chess prodigy who finished his high school exams two years early before coding the multi-million selling simulation game Theme Park at age 17. Following graduation from Cambridge University with a Double First in Computer Science he founded the pioneering videogames company Elixir Studios producing award winning games for global publishers such as Vivendi Universal. After a decade of experience leading successful technology startups, Demis returned to academia to complete a PhD in cognitive neuroscience at University College London, followed by postdoctoral research at MIT and Harvard. His research into the neural mechanisms underlying imagination and planning was listed in the top ten scientific breakthroughs of 2007 by the journal Science.

德米斯曾五次获得世界运动会冠军，是英国皇家艺术学会和皇家工程院院士，曾荣获皇家工程院银质奖章。2017 年，他被《时代》杂志评为全球最具影响力的 100 人之一，2018 年因在科学和技术方面的贡献荣获 CBE 勋章。他被选为英国皇家学会院士，曾获得该学会的穆拉德奖，还被伦敦帝国理工学院授予荣誉博士学位。

Demis is a five-time World Games Champion, and a Fellow of the Royal Society of Arts and the Royal Academy of Engineering, winning the Academy’s Silver Medal. In 2017 he was named in the Time 100 list of the world’s most influential people, and in 2018 was awarded a CBE for services to science and technology. He was elected as a Fellow of the Royal Society, has been a recipient of the Society’s Mullard Award, and was also awarded an Honorary Doctorate by Imperial College London.

2010 年，德米斯与 Shane Legg 和 Mustafa Suleyman 共同创立了 DeepMind。2014 年，DeepMind 被谷歌收购，目前是 Alphabet 旗下子公司。2016 年，DeepMind 的 AlphaGo 系统击败了李世石，后者可以说是世界围棋高手。纪录片《AlphaGo》（https://www.alphagomovie.com/）记录了这场比赛。

Demis co-founded DeepMind along with Shane Legg and Mustafa Suleyman in 2010. DeepMind was acquired by Google in 2014 and is now part of Alphabet. In 2016 DeepMind’s AlphaGo system defeated Lee Sedol, arguably the world’s best player of the ancient game of Go. That match is chronicled in the documentary film AlphaGo (https://www.alphagomovie.com/).

LANDING AI 首席执行官兼 AI FUND 普通合伙人斯坦福大学计算机科学兼职教授

CEO, LANDING AI & GENERAL PARTNER, AI FUND ADJUNCT PROFESSOR COMPUTER SCIENCE, STANFORD

吴恩达因其在人工智能和深度学习方面的贡献而广受认可，他既是学术研究人员，也是企业家。他是 Google Brain 项目和在线教育公司 Coursera 的联合创始人。随后，他成为百度的首席科学家，并在那里建立了一个行业领先的人工智能研究团队。吴恩达在谷歌和百度向人工智能驱动型组织的转型中发挥了重要作用。2018 年，他成立了 AI Fund，这是一家专注于从零开始创建人工智能领域初创公司的风险投资公司。

Andrew Ng is widely recognized for his contributions to artificial intelligence and deep learning, as both an academic researcher and an entrepreneur. He co-founded both the Google Brain project and the online education company, Coursera. He then became the chief scientist at Baidu, where he built an industry-leading AI research group. Andrew played a major role in the transformation of both Google and Baidu into AI-driven organizations. In 2018 he established AI Fund, a venture capital firm focused on building startup companies in the AI space from scratch.

马丁·福特：我们先来谈谈人工智能的未来。深度学习取得了显著的成功，但也引发了巨大的炒作。您是否认为深度学习是未来的发展方向，是人工智能继续进步的主要思想？或者从长远来看，一种全新的方法可能会取代它？

MARTIN FORD: Let’s start by talking about the future of AI. There’s been remarkable success, but also enormous hype, associated with deep learning. Do you feel that deep learning is the way forward and—the primary idea that will continue to underlie progress in AI? Or is it possible that an entirely new approach will replace it in the long run?

ANDREW NG：我真的希望有比深度学习更好的东西。最近人工智能崛起所带来的所有经济价值都归功于监督学习——基本上就是学习输入和输出映射。例如，对于自动驾驶汽车，输入是汽车前方的视频图片，输出是其他汽车的实际位置。还有其他例子，语音识别的输入是音频片段，输出是文本记录；机器翻译的输入是英文文本，输出是中文文本。

ANDREW NG: I really hope there’s something else out there better than deep learning. All of the economic value driven by this recent rise of AI is down to supervised learning—basically learning input and output mappings. For example, with self-driving cars the input is a video picture of what’s in front of your car, and the output is the actual position of the other cars. There are other examples, speech recognition has an input of an audio clip and an output of a text transcript, machine translation has an input of English text and an output of Chinese text, say.

深度学习对于学习这些输入/输出映射非常有效，这被称为监督学习，但我认为人工智能比监督学习要大得多。

Deep learning is incredibly effective for learning these input/output mappings and this is called supervised learning, but I think that artificial intelligence is much bigger than supervised learning.

监督学习的兴起为几乎所有主要行业创造了大量机会。监督学习具有极高的价值，将改变多个行业，但我认为还有很大的空间可以发明更好的东西。不过，现在很难确切地说那会是什么。

The rise of supervised learning has created a lot of opportunities in probably every major industry. Supervised learning is incredibly valuable and will transform multiple industries, but I think that there is a lot of room for something even better to be invented. It’s hard to say right now exactly what that would be, though.

马丁·福特：通向通用人工智能的道路是怎样的？您认为，要实现通用人工智能，我们必须实现哪些主要突破？

MARTIN FORD: What about the path to artificial general intelligence? What would you say are the primary breakthroughs that have to occur for us to get to AGI?

ANDREW NG：我认为这条道路非常不明确。我们可能需要的东西之一就是无监督学习。例如，今天为了教计算机什么是咖啡杯，我们向它展示了数千个咖啡杯，但无论孩子的父母多么耐心和爱心，都不会向孩子指出数千个咖啡杯。孩子们学习的方式是通过环游世界并沉浸在图像和音频中。作为孩子的经历让他们知道什么是咖啡杯。无需父母或标注者指出数千个咖啡杯，从无标记数据中学习的能力对于使我们的系统更加智能至关重要。

ANDREW NG: I think the path is very unclear. One of the things we will probably need is unsupervised learning. For example, today in order to teach a computer what a coffee mug is we show it thousands of coffee mugs, but no child’s parents, no matter how patient and loving, ever pointed out thousands of coffee mugs to that child. The way that children learn is by wandering around the world and soaking in images and audio. The experience of being a child allows them to learn what a coffee mug is. The ability to learn from unlabeled data, without parents or labelers pointing out thousands of coffee mugs, will be crucial to making our systems more intelligent.

我认为人工智能的一个问题是我们在构建专用智能或狭义智能方面取得了很大进展，而在通用人工智能方面却进展甚微。问题是，这两者都被称为人工智能。人工智能对在线广告、语音识别和自动驾驶汽车来说非常有价值，但它是专用智能，而不是通用智能。公众看到的大部分是构建专用智能的进展，他们认为我们因此在人工智能方面取得了快速进展。事实并非如此。

I think one of the problems in AI is that we’ve made a lot of progress in building specialized intelligence or narrow intelligence, and very little progress towards AGI. The problem is, both of these things are called AI. AI turns out to be incredibly valuable for online advertising, speech recognition and self-driving cars, but it’s specialized intelligence, not general. Much of what the public sees is progress in building specialized intelligence and they think that we are therefore making rapid progress toward artificial general intelligence. It’s just not true.

我很想实现 AGI，但这条路还很不明朗。我认为对 AI 了解较少的人使用了非常简单的推断，这导致了对 AI 不必要的炒作。

I would love to get to AGI, but the path is very unclear. I think that individuals that are less knowledgeable about AI have used very simplistic extrapolations, and that has led to unnecessary amounts of hype about AI.

马丁·福特：您希望在有生之年实现通用人工智能 (AGI) 吗？

MARTIN FORD: Do you expect AGI to be achieved in your lifetime?

ANDREW NG：说实话，我真的不知道。我希望在有生之年看到 AGI 的出现，但我认为它很有可能还要更久以后才会出现。

ANDREW NG: The honest answer is that I really don’t know. I would love to see AGI in my lifetime, but I think there’s a good chance it’ll be further out than that.

马丁·福特：您是如何对人工智能产生兴趣的？这又如何为您带来如此丰富的职业轨迹？

MARTIN FORD: How did you become interested in AI? And how did that lead to such a varied career trajectory?

ANDREW NG：我第一次接触神经网络是在高中时，当时我做办公室助理实习。实习和神经网络之间似乎没有明显的联系，但在实习期间，我思考过如何将我所做的一些工作自动化，那是我最早考虑神经网络的时候。我最终在卡内基梅隆大学获得学士学位，在麻省理工学院获得硕士学位，并在加州大学伯克利分校获得博士学位，论文题目为《强化学习中的塑造和策略搜索》。

ANDREW NG: My first encounter with neural networks was when I was in high school where I did an office assistant internship. There may not seem like an obvious link between an internship and neural networks, but during the course of my internship I thought about how we could automate some of the work that I was doing, and that was the earliest time I was thinking about neural networks. I wound up doing my bachelor’s at Carnegie Mellon, my master’s from MIT and a PhD, with a thesis titled, Shaping and Policy Search in Reinforcement Learning, from the University of California, Berkeley.

在接下来的大约十二年里，我在斯坦福大学计算机科学系和电气工程系担任教授。然后在 2011 年至 2012 年间，我是 Google Brain 团队的创始成员，该团队帮助 Google 转变为我们现在所认为的 AI 公司。

For about the next twelve years I taught at the Stanford University Department of Computer Science and the Department of Electrical Engineering as a professor. Then between 2011 and 2012, I was a founding member of the Google Brain team, which helped transform Google into the AI company that we now perceive it to be.

马丁·福特：谷歌大脑是谷歌首次真正使用深度学习的尝试，对吗？

MARTIN FORD: And Google Brain was the first attempt to really use deep learning at Google, correct?

ANDREW NG：在一定程度上是这样的。虽然有一些小型项目是基于神经网络的，但 Google Brain 团队才是真正将深度学习带入 Google 很多领域的推动力量。我领导 Brain 团队时做的第一件事就是在 Google 内部为大约 100 名工程师开设了一门课程。这帮助很多 Google 工程师学习了深度学习，也为 Google Brain 团队带来了很多盟友和合作伙伴，让更多人接触到了深度学习。

ANDREW NG: To an extent. There had been some small-scale projects based around neural networks, but the Google Brain team really was the force that took deep learning into many parts of Google. The first thing I did when I was leading the Brain team was to teach a class within Google for around 100 engineers. This helped teach a lot of Google engineers about deep learning, and it created a lot of allies and partners for the Google Brain team and opened up deep learning to a lot more people.

我们开展的前两个项目是与语音团队合作，我认为这有助于改变谷歌的语音识别，以及研究无监督学习，这导致​​了臭名昭著的谷歌猫的诞生。在这个项目中，我们在 YouTube 数据上免费设置了一个无监督神经网络，它学会了识别猫。无监督学习并不是今天真正创造价值的东西，但那是一个很好的技术演示，展示了我们当时使用谷歌的计算集群可以实现的规模类型。我们能够实现非常大规模的深度学习算法。

The first two projects we did were partnering with the speech team, which I think helped transform speech recognition at Google, and working on unsupervised learning, which led to the somewhat infamous Google cat. This is where we set an unsupervised neural network free on YouTube data and it learned to recognize cats. Unsupervised learning isn’t what actually creates the most value today, but that was a nice technology demonstration of the type of scale we could achieve using Google’s compute cluster at the time. We were able to do very large-scale deep learning algorithms.

马丁·福特：您在谷歌一直待到 2012 年。接下来您有什么打算呢？

MARTIN FORD: You stayed at Google until 2012. What came next for you?

ANDREW NG：在我离开谷歌的最后阶段，我觉得深度学习应该转向 GPU。因此，我最终在斯坦福大学而不是谷歌从事这项工作。事实上，我记得在 NIPS（神经信息处理系统年度会议）上与 Geoff Hinton 的一次对话，当时我试图使用 GPU，我认为这后来影响了他与 Alex Krizhevsky 的工作，并影响了很多人采用 GPU 进行深度学习。

ANDREW NG: Towards the end of my time at Google, I felt that deep learning should move toward GPUs. As a result, I wound up doing that work at Stanford University rather than at Google. In fact, I remember a conversation that I had with Geoff Hinton at NIPS, the annual conference on Neural Information Processing Systems, where I was trying to use GPUs, and I think that later influenced his work with Alex Krizhevsky and influenced, quite a lot of people to then adopt GPUs for deep learning.

我当时很幸运能在斯坦福大学任教，因为在硅谷，我们看到了通用 GPU（通用 GPU）计算即将到来的信号。我们在正确的时间出现在正确的地点，我们在斯坦福大学有朋友在研究 GPGPU，所以我们比几乎其他人更早地看到了 GPU 帮助扩展深度学习算法的能力。

I was lucky to be teaching at Stanford at the time because being here in Silicon Valley, we saw the signals that GPGPU (general-purpose GPU) computing was coming. We were in the right place at the right time and we had friends at Stanford working on GPGPUs, so we saw the ability of GPUs to help scale up deep learning algorithms earlier than almost everyone else.

我以前在斯坦福大学的学生亚当·科茨 (Adam Coates) 实际上是我决定向拉里·佩奇 (Larry Page) 推荐 Google Brain 团队的原因，目的是让拉里批准我使用他们的大量计算机来构建一个非常大的神经网络。这实际上是亚当·科茨 (Adam Coates) 生成的一个数字，其中 x 轴表示数据量，y 轴表示算法的性能。亚当生成的这个数字表明，我们训练这些深度学习算法的数据越多，它们的性能就越好。

My former student at Stanford, Adam Coates was actually the reason I decided to pitch the Google Brain team to Larry Page in a bid to get Larry to approve me using a lot of their computers to build a very large neural network. It was really one figure that was generated by Adam Coates, where the x-axis was the amount of data, and the y-axis was the performance of an algorithm. Adam generated this figure showing that the more data we could train these deep learning algorithms on, the better they’d perform.

马丁·福特：之后，你和达芙妮·科勒一起创办了Coursera，这本书也采访了她。然后你去了百度。你能描述一下你担任这些职务的历程吗？

MARTIN FORD: After that you went on to start Coursera with Daphne Koller, who is also interviewed in this book. Then you moved on to Baidu. Can you describe your path through those roles?

吴恩达：是的，我和 Daphne 一起创办了 Coursera，因为我想将 AI 和其他领域的在线教学扩展到全球数百万人。我觉得 Google Brain 团队当时已经势头强劲，所以我很高兴将权力移交给 Jeff Dean，然后转而加入 Coursera。我从头开始创建 Coursera 几年，直到 2014 年我离开那里的日常工作，前往百度的 AI 集团工作。就像 Google Brain 帮助 Google 转变为今天大家所认为的 AI 公司一样，百度 AI 集团也为将百度转变为现在很多人所认为的 AI 公司做了很多工作。在百度，我组建了一支团队，负责开发技术、支持现有业务部门，然后系统地使用 AI 开展新业务。

ANDREW NG: Yes, I helped to start Coursera with Daphne because I wanted to scale online teaching both around AI and other things to millions of people around the world. I felt that the Google Brain team already had tremendous momentum at that point, so I was very happy to hand the reins over to Jeff Dean and move on to Coursera. I worked at building Coursera from the ground up for a couple of years until 2014 when I stepped away from my day-to-day work there to go and work at Baidu’s AI Group. Just as Google Brain helped transform Google into the AI company you perceive it to be today, the Baidu AI group did a lot of work to transform Baidu into the AI company that a lot of people now perceive Baidu to be. At Baidu, I built a team that built technology, supported existing business units, and then systematically initiated new businesses using AI.

在那里工作三年后，团队运转得非常好，所以我决定再次继续前进，这次成为 Landing AI 的首席执行官和 AI Fund 的普通合伙人。

After three years there the team was running very well, so I decided to move on again this time becoming the CEO of Landing AI and a general partner at AI Fund.

马丁·福特：您在将谷歌和百度转型为人工智能驱动型公司方面发挥了重要作用，现在听起来您想扩大规模并改变其他一切。这是您对人工智能基金和 Landing AI 的愿景吗？

MARTIN FORD: You’ve been instrumental in transforming both Google and Baidu into AI-driven companies, and it sounds like now you want to scale that out and transform everything else. Is that your vision for AI Fund and Landing AI?

ANDREW NG：是的，我已经完成了大型网络搜索引擎的转型，现在我更想去转型其他行业。在 Landing AI，我帮助使用人工智能的公司转型。人工智能为现有公司提供了大量机会，因此 Landing AI 专注于帮助那些已经存在的公司转型并抓住这些人工智能机会。AI Fund 更进一步，寻找围绕人工智能技术从头开始创建的新创业公司和新企业的机会。

ANDREW NG: Yes, I’m done transforming large web search engines, and now I’d rather go and transform some other industries. At Landing AI, I help to transform companies using AI. There are a lot of opportunities in AI for incumbent companies, so Landing AI is focused on helping those companies that already exist to transform and embrace those AI opportunities. AI Fund takes this a step further, looking at the opportunities for new startups and new businesses to be created from scratch built around AI technologies.

这些模式截然不同，机遇也各异。例如，如果你看看互联网最近发生的重大技术变革，你会发现苹果和微软等老牌公司在转型成为互联网公司方面做得非常出色。然而，你只需看看谷歌、亚马逊、百度和 Facebook 等“初创公司”现在有多大，以及它们如何在互联网崛起的基础上打造出如此有价值的企业。

These are very different models with different opportunities. For example, if you look at the recent major technological transformation of the internet, incumbent companies like Apple and Microsoft did a great job transforming themselves to be internet companies. However, you only have to look at how big the “startups,” like Google, Amazon, Baidu, and Facebook are now and how they did such a great job building incredibly valuable businesses based on the rise of the internet.

随着人工智能的崛起，一些老牌公司也将出现，具有讽刺意味的是，其中许多公司都是上个时代的初创公司，如谷歌、亚马逊、Facebook 和百度，它们将在人工智能的崛起中取得巨大成功。人工智能基金正试图创建能够利用我们拥有的这些新人工智能能力的新初创公司。我们希望找到或创建下一个谷歌或 Facebook。

With the rise of AI there will also be some incumbent companies, ironically many of them were startups in the previous age, like Google, Amazon, Facebook, and Baidu, that’ll do very well with the rise of AI. AI Fund is trying to create the new startup companies that leverage these new AI capabilities we have. We want to find or create the next Google or Facebook.

马丁·福特：很多人说，谷歌和百度等巨头基本上是不可撼动的，因为他们掌握着大量数据，这为小公司进入该领域设置了障碍。您认为初创公司和小公司在人工智能领域会很难获得发展吗？

MARTIN FORD: There are a lot of people who say that the incumbents like Google and Baidu are essentially unshakable because they have access to so much data, and that creates a barrier to entry for smaller companies. Do you think startups and smaller companies are going to struggle to get traction in the AI space?

安德鲁·吴 (ANDREW NG)：大型搜索引擎所拥有的数据资产无疑为网络搜索业务设置了一道极为坚固的壁垒，但与此同时，网络搜索点击流数据对于医疗诊断、制造业或个性化教育辅导等方面的用途却并不明显。

ANDREW NG: That data asset that the large search engines have definitely creates a highly defensible barrier to the web search business, but at the same time, it’s not obvious how web search clickstream data is useful for medical diagnosis or for manufacturing or for personalized educational tutors, for example.

我认为数据实际上是垂直化的，因此，利用来自某个垂直领域的大量数据，可以在该垂直领域打造一个有防御能力的企业。正如 100 年前电力改变了多个行业一样，人工智能也将改变多个行业，我认为，多家公司都有足够的空间获得巨大成功。

I think data is actually verticalized, so building a defensible business in one vertical can be done with a lot of data from that vertical. Just as electricity transformed multiple industries 100 years ago, AI will transform multiple industries, and I think that there is plenty of room for multiple companies to be very successful.

马丁·福特：您提到了您最近创立的 AI Fund，我认为它的运作方式与其他风险投资基金不同。您对 AI Fund 的愿景是什么？它有何独特之处？

MARTIN FORD: You mentioned AI Fund, which you founded recently and which I think operates differently from other venture capital funds. What is your vision for AI Fund, and how is it unique?

ANDREW NG：是的。AI Fund 与大多数风险投资基金截然不同，我认为大多数风险投资基金的业务是寻找赢家，而我们业务是创造赢家。我们从零开始创建初创企业，我们告诉企业家，如果你已经有了演讲稿，那么你可能已经处于太晚的阶段了。

ANDREW NG: Yes. AI Fund is extremely different from most venture capital funds, and I think most venture capital funds are in the business of trying to identify winners, while we’re in the business of creating winners. We build startups from scratch, and we tell entrepreneurs that if you already have a pitch deck, you’re probably at too late a stage for us.

我们招募团队作为员工，与他们一起工作，指导他们，支持他们，尽一切努力从零开始建立一家成功的创业公司。我们实际上告诉人们，如果你有兴趣和我们一起工作，不要给我们发一份演讲稿，而是给我们一份简历，然后我们会一起努力充实创业想法。

We bring in teams as employees and work with them, mentor them, and support them, whatever is needed to try and build a successful startup from scratch. We actually tell people that if you’re interested in working with us, don’t send us a pitch deck, send us a resume and then we’ll work together to flesh out the startup idea.

马丁·福特：大多数来找您的人是不是已经有了想法，还是您会帮助他们想出办法？

MARTIN FORD: Do most people that come to you already have an idea, or do you help them come up with something?

ANDREW NG：如果他们有想法，我们很乐意讨论，但我的团队有很多想法，我们认为这些想法很有前途，但我们没有足够的资金去投资。当有人加入我们时，我们非常乐意与他们分享这一长串想法，看看哪些是合适的。

ANDREW NG: If they have an idea we’re happy to talk about it, but my team has a long list of ideas that we think are promising but we don’t have the bandwidth to invest in. When people join us, we’re very happy to share this long list of ideas with them to see which ones fit.

马丁·福特：听起来你的策略是通过提供创办初创企业的机会和基础设施来吸引人工智能人才。

MARTIN FORD: It sounds like your strategy is to attract AI talent in part by offering the opportunity and infrastructure to found a startup venture.

ANDREW NG：是的，打造一家成功的人工智能公司需要的不仅仅是人工智能人才。我们之所以如此关注技术，是因为技术发展如此迅速，但打造一支强大的人工智能团队往往需要具备多种技能，包括技术、商业战略、产品、营销和业务发展。我们的角色是打造能够构建具体业务垂直领域的全栈团队。技术非常重要，但创业公司不仅仅是技术。

ANDREW NG: Yes, building a successful AI company takes more than AI talent. We focus so much on the technology because it’s advancing so quickly, but building a strong AI team often needs a portfolio of different skills ranging from the tech, to the business strategy, to product, to marketing, to business development. Our role is building full stack teams that are able to build concrete business verticals. The technology is super important, but a startup is much more than technology.

马丁·福特：到目前为止，似乎任何表现出真正潜力的人工智能初创公司都会被大型科技公司收购。您认为最终会有人工智能初创公司上市并成为上市公司吗？

MARTIN FORD: So far, it seems that any AI startup that demonstrates real potential gets acquired by one of the huge tech firms. Do you think that eventually there’ll be AI startups that will go on to have IPOs and become public companies?

ANDREW NG：我真的希望会有很多优秀的人工智能初创公司，而不是被规模更大的初创公司收购。将首次公开募股作为一种策略并不是我们的目标，但我确实希望会有很多非常成功的人工智能初创公司，最终作为独立实体长期蓬勃发展。我们并没有真正的财务目标；我们的目标是为世界做一些好事。如果每家人工智能初创公司最终都被一家大公司收购，我会感到非常难过，我认为我们不会走向那个方向。

ANDREW NG: I really hope there’ll be plenty of great AI startups that are not just acquired by much larger startups. Initial public offering as a tactic is not the goal, but I certainly hope that there’ll be many very successful AI startups that will end up thriving as standalone entities for a long time. We don’t really have a financial goal; the goal is to do something good in the world. I’d be really sad if every AI startup ends up being acquired by a bigger company, and I don’t think we’re headed there.

马丁·福特：最近，我听到很多人表达了这样的观点：深度学习被过度炒作，可能很快就会在持续发展方面“遭遇瓶颈”。甚至有人认为，新的人工智能寒冬即将来临。您认为这是一个真正的风险吗？幻灭会导致投资大幅下降吗？

MARTIN FORD: Lately, I’ve heard a number of people express the view that deep learning is over-hyped and might soon “hit a wall” in terms of continued progress. There have even been suggestions that a new AI Winter could be on the horizon. Do you think that’s a real risk? Could disillusionment lead to a big drop off in investment?

ANDREW NG：不，我不认为会再出现一次人工智能寒冬，但我认为需要重新设定对 AGI 的期望。在早期的人工智能寒冬中，人们对最终没有真正实现的技术进行了大量的炒作。被炒作的技术实际上并没有那么有用，而早期技术创造的价值远远低于预期。我认为这就是导致人工智能寒冬的原因。

ANDREW NG: No, I don’t think there’ll be another AI winter, but I do think there needs to be a reset of expectations about AGI. In the earlier AI winters, there was a lot of hype about technologies that ultimately did not really deliver. The technologies that were hyped were really not that useful, and the amount of value created by those earlier generations of technology was vastly less than expected. I think that’s what caused the AI winters.

在当今时代，如果你看看迄今为止实际从事深度学习项目的人数，就会发现这个数字比六个月前多得多，而六个月前又比六个月前多得多。深度学习领域的具体项目数量、研究深度学习的人数、学习深度学习的人数以及基于深度学习的公司数量意味着，深度学习产生的收入实际上正在非常强劲地增长。

In the current era, if you look at the number of people actually working on deep learning projects to date, it’s much greater than six months ago, and six months ago, it was much greater than six months before that. The number of concrete projects in deep learning, the number of people researching it, the number of people learning it, and the number of companies being built on it means the amount of revenue being generated is actually growing very strongly.

经济基本面支持继续投资深度学习。大公司继续大力支持深度学习，这不仅仅是基于希望和梦想，而是基于我们已经看到的结果。这将使信心继续增长。现在，我确实认为我们需要重新设定对整个人工智能，特别是 AGI 的期望。我认为深度学习的兴起不幸地伴随着实现 AGI 的必经之路的虚假希望和梦想，我认为重新设定每个人对此的期望将非常有帮助。

The fundamentals of the economics support continued investment in deep learning. Large companies are continuing to back deep learning strongly, and it’s not based on just hopes and dreams, it’s based on the results we’re already seeing. That will see confidence continue to grow. Now, I do think we need to reset the expectations about AI as a whole, and AGI in particular. I think the rise of deep learning was unfortunately coupled with false hopes and dreams of a sure path to achieving AGI, and I think that resetting everyone’s expectations about that would be very helpful.

马丁·福特：那么，除了对 AGI 的不切实际的期望之外，您是否认为我们会继续看到深度学习在更狭窄的应用领域取得持续进展？

MARTIN FORD: So, aside from unrealistic expectations about AGI, do you think we will continue to see consistent progress with the use of deep learning in more narrow applications?

ANDREW NG：我认为当前这一代人工智能有很多局限性。不过，人工智能是一个广泛的范畴，我认为当人们讨论人工智能时，他们真正指的是反向传播、监督学习和神经网络等特定工具集。这是人们目前正在研究的最常见的深度学习部分。

ANDREW NG: I think there are a lot of limitations to the current generation of AI. AI is a broad category, though, and I think when people discuss AI, what they really mean is the specific toolset of backpropagation, supervised learning, and neural networks. That is the most common piece of deep learning that people are working on right now.

当然，深度学习是有限的，就像互联网是有限的，电力也是有限的一样。我们发明了电力作为一种公用设施，但它并没有突然解决人类的所有问题。同样，反向传播不会解决人类的所有问题，但它被证明是非常有价值的，而且我们还远没有完成用反向传播训练的神经网络可以做的所有事情。我们只是处于弄清楚当前这一代技术的影响的早期阶段。

Of course, deep learning is limited, just like the internet is limited, and electricity is limited. Just because we invented electricity as a utility, it didn’t suddenly solve all of the problems of humanity. In the same way, backpropagation will not solve all the problems of humanity, but it is turning out to be incredibly valuable, and we’re nowhere near done building out all the things we could do with neural networks trained by backpropagation. We’re just in the early phases of figuring out the implications of even the current generation of technology.

有时，当我谈论人工智能时，我首先会说“人工智能不是魔法，它不能做所有的事情。”我觉得很奇怪，我们生活在这样一个世界里，任何人都不得不说这样的话——有一种技术不能做所有的事情。

Sometimes, when I’m giving a talk about AI, the first thing I say is “AI is not magic, it can’t do everything.” I think it’s very strange that we live in a world where anyone even has to say sentences like that—that there’s a technology that cannot do everything.

人工智能面临的巨大问题就是我所说的沟通问题。狭义人工智能取得了巨大进步，通用人工智能也取得了真正的进步，但这两样东西都被称为人工智能。因此，狭义人工智能在经济和价值方面取得的巨大进步正确地让人们看到了人工智能的巨大进步，但也让人们错误地认为通用人工智能也取得了巨大的进步。坦率地说，我没有看到太多进展。除了拥有更快的计算机和数据，以及在非常普遍的层面上取得的进展外，我没有看到通用人工智能的具体进展。

The huge problem that AI has had is what I call the communications problem. There’s been tremendous progress in narrow artificial intelligence and also real progress in artificial general intelligence, but both of these things are called AI. So, tremendous progress in economics and value through narrow artificial intelligence is rightly causing people to see that there’s tremendous progress in AI, but it’s also causing people to falsely reason that there’s tremendous progress in AGI as well. Frankly, I do not see much progress. Other than having faster computers and data, and progress at a very general level, I do not see specific progress toward AGI.

马丁·福特：关于人工智能的未来，似乎存在两大阵营。一些人认为它将始终是神经网络，而另一些人则认为，要取得持续进展，需要采用一种混合方法，结合其他领域的思想，例如符号逻辑。您的看法是什么？

MARTIN FORD: There seem to be two general camps with regard to the future of AI. Some people believe it will be neural networks all the way, while others think a hybrid approach that incorporates ideas from other areas, for example symbolic logic, will be required to achieve continued progress. What’s your view?

ANDREW NG：我认为这取决于你谈论的是短期还是长期。在 Landing AI，我们一直使用混合工具为工业合作伙伴构建解决方案。通常会将深度学习工具与传统计算机视觉工具混合使用，因为当你的数据集很小的时候，深度学习本身并不总是最好的工具。成为 AI 人员的技能之一就是知道何时使用混合工具以及如何将一切结合在一起。这就是我们提供大量短期有用应用程序的方式。

ANDREW NG: I think it depends on whether you’re talking short term or long term. At Landing AI we use hybrids all the time to build solutions for industrial partners. There’s often a hybrid of deep learning tools together with, say, traditional computer vision tools because when your datasets are small, deep learning by itself isn’t always the best tool. Part of the skill of being an AI person is knowing when to use a hybrid and how to put everything together. That’s how we deliver tons of short-term useful applications.

总的来说，人们已经从传统工具转向深度学习，特别是当你拥有大量数据时，但世界上仍然有很多问题，你只有小型数据集，然后技巧在于设计混合并获得正确的技术组合。

On balance, there’s been a shift from traditional tools toward deep learning, especially when you have a lot of data, but there are still plenty of problems in the world where you have only small datasets, and then the skill is in designing the hybrid and getting the right mix of techniques.

我认为从长远来看，如果我们真的要向更接近人类水平的智能迈进，也许不是通用人工智能，而是更灵活的学习算法，我认为我们将继续看到向神经网络的转变，但尚未发明的最令人兴奋的事情之一将是比反向传播好得多的其他算法。就像交流电极其有限，但也极其有用一样，我认为反向传播也极其有限，但也极其有用，在这种情况下，我认为没有任何矛盾。

I think in the long term, if we ever move toward more human-level intelligence, maybe not for AGI but more flexible learning algorithms, I think that we’ll continue to see a shift toward neural networks, but one of the most exciting things yet to be invented will be other algorithms that are much better than backpropagation. Just like alternating current power is incredibly limited, but also incredibly useful, I think backpropagation is also incredibly limited, but incredibly useful, and I don’t see any contradiction in those circumstances.

马丁·福特：那么，就您而言，神经网络显然是推动人工智能发展的最佳技术？

MARTIN FORD: So, as far as you’re concerned, neural networks are clearly the best technology to take AI forward?

ANDREW NG：我认为在可预见的未来，神经网络将在人工智能领域占据非常重要的地位。我看不到任何可以取代神经网络的候选技术，但这并不是说未来不会出现任何技术。

ANDREW NG: I think that for the foreseeable future, neural networks will have a very central place in the AI world. I don’t see any candidates on the horizon for replacing neural networks, that’s not to say that there won’t be something on the horizon in the future.

马丁·福特：我最近与 Judea Pearl 进行了交谈，他坚信人工智能需要因果模型才能取得进步，而当前的人工智能研究并未对此给予足够的重视。您如何回应这一观点？

MARTIN FORD: I recently spoke with Judea Pearl, and he believes very strongly that AI needs a causal model in order to progress and that current AI research isn’t giving enough attention to that. How would you respond to that view?

ANDREW NG：深度学习做不到的事情有很多，因果关系就是其中之一。还有其他一些事情，比如解释能力不够好；我们需要弄清楚如何防御对抗性攻击；我们需要更好地从小数据集而不是大数据集中学习；我们需要更好地进行迁移或多任务学习；我们需要弄清楚如何更好地使用未标记数据。所以是的，反向传播有很多事情做得不好，因果关系就是其中之一。当我看到正在创建的大量高价值项目时，我并不认为因果关系是阻碍因素，但我们当然希望在这方面取得进展。我们希望在我提到的所有这些方面取得进展。

ANDREW NG: There are hundreds of different things that deep learning doesn’t do, and causality is one of them. There are other things, such as not doing explainability well enough; we need to sort out how to defend against adversarial attacks; we need to get a lot better at learning from small datasets rather than big datasets; we need to get much better at transfer or multitask learning; we need to figure out how to use unlabeled data better. So yes, there are a lot of things that backpropagation doesn’t do well, and again causality is one of them. When I look at the amount of high value projects being created, I don’t see causality as a hindering factor in them, but of course we’d love to make progress there. We’d love to make progress in all of those things I mentioned.

马丁·福特：您提到了对抗性攻击。我看到的研究表明，使用制造的数据欺骗深度学习网络相当容易。随着这项技术变得越来越普遍，这会成为一个大问题吗？

MARTIN FORD: You mentioned adversarial attacks. I’ve seen research indicating that it is fairly easy to trick deep learning networks using manufactured data. Is that going to be a big problem as this technology becomes more prevalent?

吴恩达：我认为这已经是一个问题，尤其是在反欺诈方面。当我担任百度人工智能团队负责人时，我们一直在与欺诈者作斗争，他们既攻击人工智能系统，也使用人工智能工具进行欺诈。这不是未来的事情。我现在没有打这场战争，因为我没有领导反欺诈团队，但我领导过团队，当你与欺诈作斗争时，你会感到非常对抗和零和博弈。欺诈者非常聪明，非常老练，就像我们比他们想得更远一样，他们也比他们想得更远。随着技术的发展，攻击和防御都必须发展。这是我们这些在人工智能社区中发布产品的人几年来一直在处理的事情。

ANDREW NG: I think it is already a problem, especially in anti-fraud. When I was head of the Baidu AI team we were constantly fighting against fraudsters both attacking AI systems and using AI tools to commit fraud. This is not a futuristic thing. I’m not fighting that war right now, because I’m not leading an anti-fraud team, but I have led teams and you feel very adversarial and very zero-sum when you’re fighting against fraud. The fraudsters are very smart and very sophisticated, and just as we think multiple steps ahead, they think multiple steps ahead. As the technology evolves, the attacks and the defenses will both have to evolve. This is something that those of us shipping products in the AI community have been dealing with for a few years already.

马丁·福特：隐私问题呢？尤其是在中国，面部识别技术正变得无处不在。您是否认为我们面临着人工智能被用来建立一个奥威尔式监控国家的风险？

MARTIN FORD: What about privacy issues? In China especially, facial recognition technology is becoming ubiquitous. Do you think we run the risk that AI is going to be deployed to create an Orwellian surveillance state?

吴恩达：我不是这方面的专家，所以我会听取其他人的意见。我想说​​的是，我们看到，随着技术的进步，权力可能会更加集中。我认为这在互联网上是真的，在人工智能的兴起中也是如此。越来越小的群体有可能越来越强大。权力集中可能发生在公司层面，员工相对较少的公司可以拥有更大的影响力，也可能发生在政府层面。

ANDREW NG: I’m not an expert on that, so I’ll defer to others. One thing that I would say, is that one trend we see with many rises in technology is the potential for greater concentration of power. I think this is true of the internet, and this is true again with the rise of AI. It becomes possible for smaller and smaller groups to be more and more powerful. The concentration of power can happen at the level of corporations, where corporations with relatively few employees can have a bigger influence, or at the level of governments.

少数群体掌握的技术比以往任何时候都更加强大。例如，我们已经看到的人工智能的风险之一是少数群体能够影响大量民众的投票方式，而这对民主的影响是我们需要密切关注的，以确保民主能够自我保护，使投票真正公平，代表民众的利益。我们在最近的美国大选中看到的更多是基于互联网技术而不是人工智能技术，但机会就在那里。在此之前，电视对民主和人们的投票方式产生了巨大的影响。随着技术的发展，治理和民主的性质和结构也发生了变化，这就是为什么我们必须不断更新我们的承诺，保护社会免受其滥用。

The technology available to small groups is more powerful than ever before. For example, one of the risks of AI that we have already seen is the ability of a small group to influence the way very large numbers of people vote, and the implications of that on democracy is something that we need to pay close attention to, to make sure that democracy is able to defend itself so that votes are truly fair and representative of the interests of the population. What we saw in the recent US election was based more on internet technologies rather than AI technologies, but the opportunity is there. Before that, television had a huge effect on democracy and how people voted. As technology evolves, the nature and texture of governance and democracy changes, which is why we have to constantly refresh our commitment to protecting society from its abuse.

马丁·福特：我们来谈谈人工智能最受关注的应用之一：自动驾驶汽车。它们到底还有多远的距离？想象一下，你在一座城市里，你要叫一辆完全自动驾驶的汽车，它会把你从一个随机地点带到另一个地点。你认为什么时候这项服务才能广泛应用？

MARTIN FORD: Let’s talk about one of the highest-profile applications of AI: self-driving cars. How far off are they really? Imagine you’re in a city and you’re going to call for a fully autonomous car that will take you from one random location to another. What’s the time frame for when you think that becomes a widely available service?

ANDREW NG：我认为在特定地理围栏区域内实现自动驾驶汽车将会很快实现，可能在今年年底，但要在更普遍的环境下实现自动驾驶汽车则还需要很长时间，可能还要几十年的时间。

ANDREW NG: I think that self-driving cars in geofenced regions will come relatively soon, possibly by the end of this year, but that self-driving cars in more general circumstances will be a long way off, possibly multiple decades.

马丁·福特：你所说的地理围栏是指基本上在虚拟有轨电车轨道上运行的自动驾驶汽车，还是换句话说，只在经过深入绘制的路线上运行？

MARTIN FORD: By geofenced, you mean autonomous cars that are running essentially on virtual trolley tracks, or in other words only on routes that have been intensively mapped?

ANDREW NG：没错！不久前，我与他人合作撰写了一篇 Wired 文章，讨论了 Train Terrain（ https://www.wired.com/2016/03/self-driving-cars-wont-work-change-roads-attitudes/），其中谈到了我对自动驾驶汽车可能如何推广的看法。在大规模采用自动驾驶汽车之前，我们需要基础设施变革、社会和法律变革。

ANDREW NG: Exactly! A while back I co-authored a Wired article talking about Train Terrain (https://www.wired.com/2016/03/self-driving-cars-wont-work-change-roads-attitudes/) about how I think self-driving cars might roll out. We’ll need infrastructure changes, and societal and legal changes, before we’ll see mass adoption of self-driving cars.

我有幸见证了自动驾驶行业 20 多年的发展。20 世纪 90 年代末，我在卡内基梅隆大学读本科时，曾与 Dean Pomerleau 一起上过一门课，研究他们的自动驾驶汽车项目，该项目根据输入的视频图像来操控汽车。这项技术很棒，但还未为时代做好准备。然后在斯坦福大学，我参加了 2007 年的 DARPA 城市挑战赛，只是其中的一小部分。

I have been fortunate to have seen the self-driving industry evolve for over 20 years now. As an undergraduate at Carnegie Mellon in the late ‘90s, I did a class with Dean Pomerleau working on their autonomous car project that steered the vehicle based an input video image. The technology was great, but it wasn’t ready for its time. Then at Stanford, I was a peripheral part of the DARPA Urban Challenge in 2007.

我们飞到维克多维尔，这是我第一次在同一个地方看到这么多的自动驾驶汽车。整个斯坦福团队在前五分钟都着迷了，看着这些汽车在没有司机的情况下飞驰而过，令人惊讶的是，五分钟后，我们就适应了，然后转身背对着它。我们只是互相聊天，而自动驾驶汽车在我们 10 米外飞驰而过，我们并没有注意。人类最了不起的一点就是我们适应新技术的速度有多快，我觉得不久之后，自动驾驶汽车就不再被称为自动驾驶汽车，而只是被称为汽车。

We flew down to Victorville, and it was the first time I saw so many self-driving cars in the same place. The whole Stanford team were all fascinated for the first five minutes, watching all these cars zip around without drivers, and the surprising thing was that after five minutes, we acclimatized to it, and we turned our backs to it. We just chatted with each other while self-driving cars zipped passed us 10 meters away, and we weren’t paying attention. One thing that’s remarkable about humanity is how quickly we acclimatize to new technologies, and I feel that it’s not going to be too long before self-driving cars are no longer called self-driving cars, they’re just called cars.

马丁·福特：我知道您是自动驾驶汽车公司 Drive.ai 的董事会成员。您能预估一下他们​​的技术什么时候会得到广泛应用吗？

MARTIN FORD: I know you’re on the board of directors of the self-driving car company Drive.ai. Do you have an estimate for when their technology will be in general use?

安德鲁·吴：他们现在正在德克萨斯州开车。让我们看看，现在几点了？有人刚打完车就去吃午饭了。重要的是这是多么平常的事。有人只是出去吃午饭，就像平常一样，他们开着自动驾驶汽车去吃午饭。

ANDREW NG: They’re driving round in Texas right now. Let’s see, what time is it? Someone’s just taken one and gone for lunch. The important thing is how mundane that is. Someone’s just gone out for lunch, like any normal day, and they’ve done it by getting in a self-driving car.

马丁·福特：你对目前自动驾驶汽车的进展有什么感想？它与你的预期相比如何？

MARTIN FORD: How do you feel about the progress you’ve seen in self-driving cars so far? How has it compared with your expectations?

ANDREW NG：我不喜欢炒作，我觉得有些公司公开表示，他们制定的自动驾驶汽车普及时间表在我看来是不切实际的。我认为自动驾驶汽车将改变交通，让人类生活更加美好。但是，我认为每个人都有一份切合实际的自动驾驶汽车路线图，比让首席执行官站在台上宣布不切实际的时间表要好得多。我认为自动驾驶世界正在努力制定更切合实际的计划，将这项技术推向市场，我认为这是一件非常好的事情。

ANDREW NG: I don’t like hype, and I feel like a few companies have spoken publicly and described what I think of as unrealistic timelines about the adoption of self-driving cars. I think that self-driving cars will change transportation, and will make human life much better. However, I think that everyone having a realistic roadmap to self-driving cars is much better than having CEOs stand on stage and proclaim unrealistic timelines. I think the self-driving world is working toward more realistic programs for bringing the tech to market, and I think that’s a very good thing.

马丁·福特：您如何看待政府监管在自动驾驶汽车和人工智能领域所发挥的作用？

MARTIN FORD: How do you feel about the role of government regulation, both for self-driving cars and AI more generally?

ANDREW NG：出于安全考虑，汽车行业一直受到严格监管。我认为，鉴于人工智能和自动驾驶汽车的发展，交通运输监管需要重新考虑。监管更周到的国家将更快地实现人工智能驱动的医疗保健系统、自动驾驶汽车或人工智能驱动的教育系统等带来的可能性，而我认为监管不够周到的国家可能会落后。

ANDREW NG: The automotive industry has always been heavily regulated because of safety, and I think that the regulation of transportation needs to be rethought in light of AI and self-driving cars. Countries with more thoughtful regulation will advance faster to embrace the possibilities enabled by, for example, AI-driven healthcare systems, self-driving cars, or AI-driven educational systems, and I think countries that are less thoughtful about regulation will risk falling behind.

监管应该针对这些特定的垂直行业，因为我们可以就结果进行良好的辩论。我们可以更轻松地定义我们想做什么和不想发生什么。我发现对人工智能进行广泛监管没那么有用。我认为，在监管特定垂直行业时，仔细考虑人工智能的影响不仅有助于垂直行业的发展，而且还有助于人工智能开发正确的解决方案，并在垂直行业中更快地被采用。

Regulation should be in these specific industry verticals because we can have a good debate about the outcomes. We can more easily define what we do and do not want to happen. I find it less useful to regulate AI broadly. I think that the act of thinking through the impact of AI in specific verticals for regulation will not only help the verticals grow but will also help AI develop the right solutions and be adopted faster across verticals.

我认为自动驾驶汽车只是政府这个更广泛主题的一个缩影。每当有技术突破时，监管机构就必须采取行动。监管机构必须采取行动，确保民主得到捍卫，即使在互联网时代和人工智能时代也是如此。除了捍卫民主，政府还必须采取行动，确保其国家为人工智能的崛起做好准备。

I think self-driving cars are only a microcosm of a broader theme here, which is the government. Every time there is a technological breakthrough, regulators must act. Regulators have to act to make sure that democracy is defended, even in the era of the internet and the era of artificial intelligence. In addition to defending democracy, governments must act to make sure that their countries are well positioned for the rise of AI.

假设政府的主要职责之一是保障公民的福祉，我认为明智的政府可以帮助国家驾驭人工智能的崛起，为人民带来更好的结果。事实上，即使在今天，一些政府使用互联网的能力也远远优于其他政府。这既涉及面向公民的外部网站和服务，也涉及内部网站和服务，比如，你的政府 IT 服务是如何组织的？

Assuming that one of governments’ primary responsibilities is the well-being of their citizens, I think that governments that act wisely can help their nations ride the rise of AI, to much better outcomes for their people. In fact, even today, some governments use the internet much better than other governments. This is about external websites and services to citizens, as well as internal ones, in terms of, how are your government IT services organized?

新加坡拥有综合医疗保健系统，每个患者都有唯一的患者 ID，这使得医疗记录的整合方式令许多其他国家羡慕不已。新加坡是一个小国，所以新加坡可能比大国更容易做到这一点，但新加坡政府改变医疗保健系统以更好地利用互联网的方式，对医疗保健系统和新加坡公民的健康产生了巨大影响。

Singapore has an integrated healthcare system, where every patient has a unique patient ID, and this allows for the integration of healthcare records in a way that is the envy of many other nations. Now, Singapore’s a small country, so maybe it’s easier for Singapore than a larger country, but the way the Singapore government has shifted the healthcare system to use the internet better, has a huge impact on the healthcare system, and on the health of the Singaporean citizens.

马丁·福特：听起来您认为政府和人工智能之间的关系应该超越仅仅监管技术。

MARTIN FORD: It sounds like you think the relationship between government and AI should extend beyond just regulating the technology.

ANDREW NG：我认为政府在人工智能的崛起中扮演着重要角色，首先要确保利用人工智能做好治理工作。例如，我们是否应该使用人工智能更好地分配政府人员？林业资源呢？我们能否使用人工智能更好地分配林业资源？人工智能能否帮助我们制定更好的经济政策？政府能否利用人工智能更好、更有效地清除欺诈行为（比如税务欺诈）？我认为人工智能将在治理领域拥有数百种应用，就像人工智能在大型人工智能公司中拥有数百种应用一样。政府应该充分利用人工智能为自己服务。

ANDREW NG: I think governments have a huge role to play in the rise of AI and in making sure that first, governance is done well with AI. For instance, should we better allocate government personnel using AI? How about the forestry resources, can we allocate that better using AI? Can AI help us set better economic policies? Can the government weed out fraud—maybe tax fraud—better and more efficiently using AI? I think AI will have hundreds of applications in governance, just as AI has hundreds of applications in the big AI companies. Governments should use AI well for themselves.

对于生态系统而言，我认为公私合作也将加速国内产业的发展，制定周到的自动驾驶汽车监管规定的政府将看到自动驾驶汽车在其社区加速发展。我非常热爱我的家乡加利福尼亚州，但加利福尼亚州的法规不允许自动驾驶汽车公司做某些事情，这就是为什么许多自动驾驶汽车公司无法将总部设在加利福尼亚州，现在几乎被迫在加利福尼亚州以外的地方运营。

For the ecosystem as well, I think public-private partnerships will accelerate the growth of domestic industry, and governments that make thoughtful regulation about self-driving cars will see self-driving accelerate in their communities. I’m very committed to my home state of California, but California regulations do not allow self-driving car companies to do certain things, which is why many self-driving car companies can’t have their home bases in California and are now almost forced to operate outside of California.

我认为，无论是在州一级还是在国家一级，那些在自动驾驶汽车、无人机以及支付系统和医疗保健系统中采用人工智能方面制定了周到政策的国家——那些在所有这些垂直领域都制定了周到政策的国家将看到这些令人惊叹的新工具如何更快地应用于解决其公民面临的一些最重要的问题。除了监管和公私合作之外，为了加速采用这些令人惊叹的工具，我认为政府还需要在教育和就业问题上提出解决方案。

I think that both at the state level as well as at the nation level, countries that have thoughtful policies about self-driving cars, about drones, and about the adoption of AI in payment systems and in healthcare systems, for example—those countries with thoughtful policies in all of these verticals will see much faster progress in how these amazing new tools can be brought to bear on some of the most important problems for their citizens. Beyond regulation and public-private partnership, to accelerate the adoption of these amazing tools, I think governments also need to come up with solutions in education and on the jobs issue.

马丁·福特：就业和经济受到的影响是我写过很多文章的领域。您是否认为我们可能正面临大规模混乱，并可能导致大面积失业？

MARTIN FORD: The impact on jobs and the economy is an area that I’ve written about a lot. Do you think we may be on the brink of a massive disruption that could result in widespread job losses?

ANDREW NG：是的，我认为这是人工智能面临的最大道德问题。虽然这项技术在社会某些领域创造财富方面非常出色，但坦率地说，我们已经把美国大部分地区和世界大部分地区抛在了后面。如果我们想创造的不仅仅是一个富裕的社会，而是一个公平的社会，那么我们还有很多重要的工作要做。坦率地说，这也是我仍然热衷于在线教育的原因之一。

ANDREW NG: Yes, and I think it’s the biggest ethical problem facing AI. Whilst the technology is very good at creating wealth in some segments of society, we have frankly left large parts of the United States and also large parts of the world behind. If we want to create not just a wealthy society but a fair one, then we still have a lot of important work to do. Frankly, that’s one of the reasons why I remain very engaged in online education.

我认为，我们的世界非常善于奖励在特定时期拥有所需技能的人。如果我们能够教育人们重新学习技能，即使他们的工作被技术取代，那么我们就有更好的机会确保下一波财富创造最终以更公平的方式分配。关于邪恶的人工智能杀手机器人的大量炒作分散了领导者的注意力，使他们无法专注于更难但更重要的对话，即我们如何解决就业问题。

I think our world is pretty good at rewarding people who have the required skills at a particular time. If we can educate people to reskill even as their jobs are displaced by technology, then we have a much better chance of making sure that this next wave of wealth creation ends up being distributed in a more equitable way. A lot of the hype about evil AI killer robots distracts leaders from the much harder, but much more important conversation about what we do about jobs.

马丁·福特：您认为全民基本收入可以解决这个问题吗？

MARTIN FORD: What do you think of a universal basic income as part of a solution to that problem?

安德鲁·吴：我不支持全民基本收入，但我认为有条件基本收入是一个更好的主意。这关乎工作尊严，我实际上支持有条件基本收入，即失业者可以获得学费。这将增加失业者获得重返劳动力市场所需技能的可能性，并为支付有条件基本收入的税基做出贡献。

ANDREW NG: I don’t support a universal basic income, but I do think a conditional basic income is a much better idea. There’s a lot about the dignity of work and I actually favor a conditional basic income in which unemployed individuals can be paid to study. This would increase the odds that someone that’s unemployed will gain the skills they need to re-enter the workforce and contribute back to the tax base that is paying for the conditional basic income.

我认为，在当今世界，零工经济中有很多工作，你可以赚到足够的工资来维持生计，但没有足够的空间来提升自己或家人的生活。我非常担心无条件的基本收入会导致更多人陷入这种低工资、低技能的工作。

I think in today’s world, there are a lot of jobs in the gig economy, where you can earn enough of a wage to get by, but there isn’t much room for lifting up yourself or your family. I am very concerned about an unconditional basic income causing a greater proportion of the human population to become trapped doing this low-wage, low-skilled work.

有条件基本收入鼓励人们不断学习，这将使许多个人和家庭生活得更好，因为我们正在帮助人们获得所需的培训，以便从事更高价值和更高收入的工作。我们看到经济学家撰写的报告中包含这样的统计数据：“20 年后，50% 的工作面临自动化的风险”，这确实很可怕，但另一方面，另外 50% 的工作没有面临自动化的风险。

A conditional basic income that encourages people to keep learning and keep studying will make many individuals and families better off because we’re helping people get the training they need to then do higher-value and better-paying jobs. We see economists write reports with statistics like “in 20 years, 50% of jobs are at risk of automation,” and that’s really scary, but the flip side is that the other 50% of jobs are not at risk of automation.

事实上，我们无法找到足够的人手来完成其中的一些工作。我们无法找到足够的医护人员，我们无法在美国找到足够的教师，令人惊讶的是，我们似乎找不到足够的风力涡轮机技术人员。

In fact, we can’t find enough people to do some of these jobs. We can’t find enough healthcare workers, we can’t find enough teachers in the United States, and surprisingly we can’t seem to find enough wind turbine technicians.

问题是，那些失业的人如何从事这些我们找不到足够人手从事的高薪、高价值的工作？答案不是让每个人都学习编程。是的，我认为很多人应该学习编程，但我们还需要在医疗保健、教育、风力涡轮机技术人员和其他需求不断增长的工作领域提高更多人的技能。

The question is, how do people whose jobs are displaced take on these other great-paying, very valuable jobs that we just can’t find enough people to do? The answer is not for everyone to learn to program. Yes, I think a lot of people should learn to program, but we also need to skill up more people in those areas of healthcare, education, and wind turbine technicians, and other in-demand rising categories of jobs.

我认为我们正在远离那种一生只能从事一种职业的世界。技术变化如此之快，有些人上大学时以为自己在做一件事，但后来他们意识到自己 17 岁时所选择的职业已经不再可行，他们应该尝试不同的职业。

I think we’re moving away from a world where you have one career in your lifetime. Technology changes so fast that there will be people that thought they were doing one thing when they went to college that will realize that the career they set out toward when they were 17-years-old is no longer viable, and that they should branch into a different career.

我们已经看到千禧一代更有可能跳槽，比如从一家公司的产品经理跳槽到另一家公司的产品经理。我认为，在未来，我们会看到越来越多的人从一家公司的材料科学家跳槽到另一家公司的生物学家，再跳槽到第三家公司的安全研究员。这不会一夜之间发生，需要很长时间才能改变。不过，有趣的是，在我的深度学习世界里，我已经看到许多人从事深度学习，他们不是主修计算机科学，而是学习物理、天文学或纯数学等学科。

We’ve seen how millennials are more likely to hop among jobs, where you go from being a product manager in one company to the product manager of a different company. I think that in the future, increasingly we’ll see people going from being a material scientist in one company to being a biologist in a different company, to being a security researcher in a third company. This won’t happen overnight, it will take a long time to change. Interestingly, though, in my world of deep learning, I already see many people doing deep learning that did not major in computer science, they did subjects like physics, astronomy, or pure mathematics.

马丁·福特：对于有志于从事人工智能或深度学习事业的年轻人，您有什么特别的建议吗？他们应该完全专注于计算机科学吗？还是脑科学或人类认知研究也很重要？

MARTIN FORD: Is there any particular advice you’d give to a young person who is interested in a career in AI, or in deep learning specifically? Should they focus entirely on computer science or is brain science, or the study of cognition in humans also important?

ANDREW NG：我会建议学习计算机科学、机器学习和深度学习。脑科学或物理学的知识都很有用，但进入人工智能领域最省时的方法还是学习计算机科学、机器学习和深度学习。我认为，有了 YouTube 视频、演讲和书籍，人们可以比以往更轻松地找到资料并自行学习，只需循序渐进即可。事情不会一蹴而就，但我认为，循序渐进，几乎任何人都有可能成为人工智能领域的佼佼者。

ANDREW NG: I would say to study computer science, machine learning, and deep learning. Knowledge of brain science or physics is all useful, but the most time-efficient route to a career in AI is computer science, machine learning and deep learning. Because of YouTube videos, talks, and books, I think it’s easier than ever for someone to find materials and study by themselves, just step by step. Things don’t happen overnight, but step by step, I think it’s possible for almost anyone to become great at AI.

我倾向于给人们一些建议。首先，人们不喜欢听到要掌握一个新领域需要努力工作，但这确实需要努力工作，愿意努力工作的人会学得更快。我知道每个人每周学习一定时间是不可能的，但能找到更多时间学习的人会学得更快。

There are a couple of pieces of advice that I tend to give to people. Firstly, people don’t like to hear that it takes hard work to master a new field, but it does take hard work, and the people who are willing to work hard at it will learn faster. I know that it’s not possible for everyone to learn a certain number of hours every week, but people that are able to find more time to study will just learn faster.

我倾向于给人们的另一个建议是，假设你现在是一名医生，并且想要进入人工智能领域——作为一名医生，你将拥有独特的优势，可以在医疗保健领域从事非常有价值的工作，而很少有人能做到这一点。如果你现在是一名物理学家，看看是否有一些关于将人工智能应用于物理学的想法。如果你是一家图书出版商，看看在图书出版方面是否有一些你可以用人工智能来做的工作，因为这是利用你的独特优势并用人工智能来补充优势的一种方式，而不是与刚进入人工智能领域的大学毕业生在更公平的环境中竞争。

The other piece of advice I tend to give people is that let’s say you’re currently a doctor and you want to break into AI—as a doctor you’d be uniquely positioned to do very valuable work in healthcare that very few others can do. If you are currently a physicist, see if there are some ideas on AI applied to physics. If you’re a book publisher, see if there’s some work you can do with AI in book publishing, because that’s one way to leverage your unique strengths and to complement that with AI, rather than competing on a more even playing field with the fresh college grad stepping into AI.

马丁·福特：除了可能对就业产生影响之外，您认为我们现在或在不久的将来还应该关注与人工智能相关的哪些风险？

MARTIN FORD: Beyond the possible impact on jobs, what are the other risks associated with AI that you think we should be concerned about now or in the relatively near future?

ANDREW NG：我喜欢将人工智能与电力联系起来。电力非常强大，通常可以用于造福人类，但也可以用于伤害人类。人工智能也一样。最终，这取决于个人、公司和政府，他们必须努力确保以积极和合乎道德的方式使用这种新的超级力量。

ANDREW NG: I like to relate AI to electricity. Electricity is incredibly powerful and on average has been used for tremendous good, but it can also be used to harm people. AI is the same. In the end, it’s up to individuals, as well as companies and governments, to try to make sure we use this new superpower in positive and ethical ways.

我认为人工智能中的偏见是另一个主要问题。从人类生成的文本数据中学习的人工智能可以了解健康、性别和种族刻板印象。人工智能团队意识到了这一点，并正在积极致力于此，我感到非常鼓舞，今天我们在减少人工智能偏见方面的想法比减少人类偏见更好。

I think that bias in AI is another major issue. AI that learns from human-generated text data can pick up on health, gender, and racial stereotypes. AI teams are aware of this and are actively working on this, and I am very encouraged that today we have better ideas for reducing bias in AI than we do for reducing bias in humans.

马丁·福特：解决人们的偏见非常困难，因此看起来用软件来解决它可能是一个更容易的问题。

MARTIN FORD: Addressing bias in people is very difficult, so it does seem like it might be an easier problem to solve in software.

ANDREW NG：是的，你可以在 AI 软件中将数字归零，这样它就会表现出更少的性别偏见，但我们没有类似有效的方法来减少人类的性别偏见。我认为，我们很快就会看到 AI 系统的偏见会比许多人类更少。这并不是说我们应该满足于偏见减少，还有很多工作要做，我们应该继续努力减少这种偏见。

ANDREW NG: Yes, you can zero a number in an AI piece of software and it will exhibit much less gender bias, we don’t have similarly effective ways of reducing gender bias in people. I think that soon we might see that AI systems will be less biased than many humans. That is not to say that we should be satisfied with just having less bias, there’s still a lot of work to do and we should keep on working to reduce that bias.

马丁·福特：那么，人们担心超级智能系统有朝一日会摆脱我们的控制，对人类构成真正的威胁吗？

MARTIN FORD: What about the concern that a superintelligent system might someday break free of our control and pose a genuine threat to humanity?

安德鲁·吴：我之前说过，如今担心 AGI 邪恶杀手机器人就像担心火星人口过剩一样。我希望一个世纪后我们能殖民火星。到那时，火星可能已经人口过剩、污染严重，甚至可能有孩子死于污染。这并不是说我冷酷无情，不在乎那些垂死的孩子——我很想找到解决方案，但我们甚至还没有登陆火星，所以我觉得很难有效地解决这个问题。

ANDREW NG: I’ve said before that worrying about AGI evil killer robots today is like worrying about overpopulation on the planet Mars. A century from now I hope that we will have colonized the planet Mars. By that time, it may well be overpopulated and polluted, and we might even have children dying on Mars from pollution. It’s not that I’m heartless and don’t care about those dying children—I would love to find a solution to that, but we haven’t even landed on the planet yet, so I find it difficult to productively work on that problem.

马丁·福特：那么您不认为，人们所说的“快速起飞”情景存在任何现实的恐惧吗？在这种情景中，AGI 系统会经历一个递归的自我完善周期，并迅速变得超级智能？

MARTIN FORD: You don’t think then that there’s any realistic fear of what people call the “fast takeoff” scenario, where an AGI system goes through a recursive self-improvement cycle and rapidly becomes superintelligent?

ANDREW NG：很多关于超级智能和指数增长的炒作都是基于非常幼稚和过于简单的推断。几乎任何事情都很容易被炒作。我不认为超级智能突然出现并转瞬即逝的风险很大，就像我不认为火星会在一夜之间变得人口过剩一样。

ANDREW NG: A lot of the hype about superintelligence and exponential growth were based on very naive and very simplistic extrapolations. It’s easy to hype almost anything. I don’t think that there is a significant risk of superintelligence coming out of nowhere and it happening in a blink of an eye, in the same way that I don’t see Mars becoming overpopulated overnight.

马丁·福特：那与中国的竞争问题呢？人们经常指出，中国具有某些优势，比如由于人口众多而可以获得更多数据，而且对隐私的担忧较少。他们会在人工智能研究方面超越我们吗？

MARTIN FORD: What about the question of competition with China? It’s often pointed out that China has certain advantages, like access to more data due to a larger population and fewer concerns about privacy. Are they going to outrun us in AI research?

安德鲁·吴：电力竞争如何展开？一些国家（如美国）的电网比一些发展中经济体的电网要强大得多，这对美国来说是个好消息。然而，我认为全球人工智能竞赛并不像大众媒体有时所报道的那样激烈。人工智能是一种令人惊叹的能力，我认为每个国家都应该弄清楚如何利用这种新能力，但我认为这场竞赛并不像大众媒体所报道的那样激烈。

ANDREW NG: How did the competition for electricity play out? Some countries like the United States have a much more robust electrical grid than some developing economies, so that’s great for the United States. However, I think the global AI race is much less of a race than the popular press sometimes presents it to be. AI is an amazing capability, and I think every country should figure out what to do with this new capability, but I think that it is much less of a race than the popular press suggests.

马丁·福特：人工智能确实有军事用途，而且可能用于制造自动化武器。联合国目前正在讨论禁止全自动武器，因此人们显然对此感到担忧。这不是未来 AGI 相关的东西，而是我们很快就会看到的东西。我们应该担心吗？

MARTIN FORD: AI clearly does have military applications, though, and potentially could be used to create automated weapons. There’s currently a debate in the United Nations about banning fully autonomous weapons, so it’s clearly something people are concerned about. That’s not futuristic AGI-related stuff, but rather something we could see quite soon. Should we be worried?

安德鲁·吴：内燃机、电力和集成电路都创造了巨大的好处，但它们都对军事有用。任何新技术都是如此，包括人工智能。

ANDREW NG: The internal combustion engine, electricity, and integrated circuits all created tremendous good, but they were all useful for the military. It’s the same with any new technology, including AI.

马丁·福特：您显然对人工智能持乐观态度。我想您相信，随着人工智能的发展，其好处将大于风险？

MARTIN FORD: You’re clearly an optimist where AI is concerned. I assume you believe that the benefits are going to outweigh the risks as artificial intelligence advances?

安德鲁·吴：是的，我很喜欢。过去几年，我有幸站在人工智能产品的前线，亲眼目睹了更出色的语音识别、更出色的网络搜索和更优化的物流网络如何帮助人们。

ANDREW NG: Yes, I do. I’ve been fortunate to be on the front lines, shipping AI products for the last several years and I’ve seen firsthand the way that better speech recognition, better web search, and better optimized logistics networks help people.

这就是我看待世界的方式，也许这是一种非常幼稚的方式。世界变得非常复杂，而且世界并不是我想要的样子。坦率地说，我很怀念以前能听政治领袖和商界领袖讲话的时光，而且能更多地相信他们说的话。

This is the way that I think about the world, which may be a very naïve way. The world’s gotten really complicated, and the world’s not the way I want it to be. Frankly, I miss the times when I could listen to political leaders and business leaders, and take much more of what they said at face value.

我怀念以前对许多公司和领导人更有信心的时光，他们行为合乎道德，言出必行。如果你想想你尚未出生的孙子或未出生的曾曾孙，我认为这个世界还没有成为你希望他们成长的地方。我希望民主能够更好地发挥作用，我希望世界更加公平。我希望更多的人行为合乎道德，考虑对其他人的实际影响，我希望世界更加公平，让每个人都有机会接受教育。我希望人们努力工作，但要努力工作并不断学习，做他们认为有意义的工作，我认为世界上许多地方还没有成为我们所希望的样子。

I miss the times when I had greater confidence in many companies and leaders to behave in an ethical way and to mean what they say and say what they mean. If you think about your as-yet-unborn grandchildren or your unborn great-great-grandchildren, I don’t think the world is yet the way that you want it to be for them to grow up in. I want democracy to work better, and I want the world to be fairer. I want more people to behave ethically and to think about the actual impact on other people, and I want the world to be fairer, for everyone to have access to and gain an education. I want people to work hard, but to work hard and to keep studying, and to do work that they find meaningful, and I think many parts of the world are not yet the way I think we would all like it to be.

每次出现技术突破，都会给我们带来改变的机会。我希望我的团队以及世界各地的其他人能够努力让世界变得更美好，就像我们想要的那样。我知道这听起来像是一个梦想家，但这正是我真正想做的。

Every time there’s a technological disruption, it gives us the opportunity to make a change. I would like my teams, as well as other people around the world to take a shot at making the world a better place in the ways that we want it to be. I know that sounds like I’m a dreamer, but that’s what I actually want to do.

马丁·福特：我认为这是一个伟大的愿景。我想问题在于，这是整个社会做出的决定，让我们走上这种乐观的未来之路。您是否有信心我们会做出正确的选择？

MARTIN FORD: I think that’s a great vision. I guess the problem is that it’s a decision for society as a whole to set us on the path to that kind of optimistic future. Are you confident that we’ll make the right choices?

安德鲁·吴 (Andrew NG)：我不认为这会是一条直线，但我认为世界上有足够多的诚实、道德和善意的人，他们有足够的机会实现这一目标。

ANDREW NG: I don’t think it will be in a straight line, but I think there are enough honest, ethical, and well-meaning people in the world to have a very good shot at it.

吴恩达 是人工智能和机器学习领域最知名的人物之一。他是谷歌大脑深度学习项目以及在线教育公司Coursera的联合创始人。2014 年至 2017 年期间，他担任百度副总裁兼首席科学家，将公司的人工智能团队打造成了一个拥有数千人的组织。他被普遍认为在谷歌和百度向人工智能驱动型公司转型过程中发挥了重要作用。

ANDREW NG is one of the most recognizable names in AI and machine learning. He co-founded the Google Brain deep learning project as well as the online education company Coursera. Between 2014 and 2017, he was a vice president and chief scientist at Baidu, where he built the company’s AI group into an organization with several thousand people. He is generally credited with playing a major role in the transformation of both Google and Baidu into AI-driven companies.

离开百度后，安德鲁接手了许多项目，包括推出旨在培养深度学习专家的在线教育平台 deeplearning.ai，以及旨在利用人工智能改造企业的 Landing AI。他目前是专注于人工智能心理健康应用的初创公司 Woebot 的董事长，也是自动驾驶汽车公司 Drive.ai 的董事会成员。他还是 AI Fund 的创始人和普通合伙人，这是一家从零开始打造新人工智能初创公司的风险投资公司。

Since leaving Baidu, Andrew has undertaken a number of projects including launching deeplearning.ai, an online education platform geared toward educating deep learning experts, as well as Landing AI, which seeks to transform enterprises with AI. He’s currently the chairman of Woebot, a startup focused on mental health applications for AI and is on the board of directors of self-driving car company Drive.ai. He is also the founder and General Partner at AI Fund, a venture capital firm that builds new AI startups from the ground up.

Andrew 现为斯坦福大学兼职教授，曾任斯坦福大学副教授兼人工智能实验室主任。他在卡内基梅隆大学获得计算机科学学士学位，在麻省理工学院获得硕士学位，在加州大学伯克利分校获得博士学位。

Andrew is currently an adjunct professor, and formerly the associate professor and Director of the AI Lab at Stanford University. He received his undergraduate degree in computer science from Carnegie Mellon University, his master’s degree from MIT, and his PhD from The University of California, Berkeley.

AFFECTIVA 首席执行官兼联合创始人

CEO & CO-FOUNDER OF AFFECTIVA

Rana el Kaliouby 是 Affectiva 的联合创始人兼首席执行官，Affectiva 是一家初创公司，专门研究感知和理解人类情感的人工智能系统。Affectiva 正在开发尖端人工智能技术，应用机器学习、深度学习和数据科学，为人工智能带来新的情商水平。Rana 积极参与关注人工智能道德问题和监管的国际论坛，以帮助确保该技术对社会产生积极影响。2017 年，她被世界经济论坛评选为全球青年领袖。

Rana el Kaliouby is the co-founder and CEO of Affectiva, a startup company that specializes in AI systems that sense and understand human emotions. Affectiva is developing cutting-edge AI technologies that apply machine learning, deep learning, and data science to bring new levels of emotional intelligence to AI. Rana is an active participant in international forums that focus on ethical issues and the regulation of AI to help ensure the technology has a positive impact on society. She was selected as a Young Global Leader by the World Economic Forum in 2017.

马丁·福特：首先我想了解一下您的背景；我特别感兴趣的是您是如何涉足人工智能领域的，以及您是如何从学术背景走到今天在 Affectiva 公司发展的。

MARTIN FORD: I want to begin by exploring your background; I’m especially interested in how you became involved with AI and how your trajectory went from an academic background to where you are today with your company, Affectiva.

RANA EL KALIOUBY：我在中东长大，出生在埃及开罗，童年的大部分时间在科威特度过。在那段时间里，我开始尝试早期的计算机，因为我的父母都是从事技术行业的，我爸爸会把旧的 Atari 机器带回家，我们会把它们拆开。很快，我就在开罗美国大学攻读了本科课程，主修计算机科学。我想你可以说，这就是 Affectiva 背后的思想第一次发挥作用的地方。在那段时间里，我开始着迷于技术如何改变人类彼此之间的联系方式。如今，我们的许多交流都是通过技术进行的，因此，我们与技术以及彼此之间的特殊联系方式让我着迷。

RANA EL KALIOUBY: I grew up around the Middle East, being born in Cairo, Egypt and spending much of my childhood in Kuwait. During this time, I found myself experimenting with early computers, as a result of both my parents being in technology, and my dad would bring home the old Atari machines where we would pick them apart. Fast-forward and that grew into my undergraduate course where I majored in Computer Science at the American University in Cairo. I guess you could say this is where the thinking behind Affectiva first came into play. During this time, I became fascinated by how technology changes how humans connect with one another. Nowadays a lot of our communication is mediated via technology, and so the special way that we connect with technology, but also with one another, fascinates me.

下一步是攻读博士学位。我获得了剑桥大学计算机科学系的奖学金，顺便说一句，这对于埃及穆斯林年轻女性来说是一件相当不寻常的事情。那是在 2000 年，那时我们还没有智能手机，但当时我对人机交互的概念以及我们的界面在未来几年将如何发展非常感兴趣。

The next step was to do a PhD. I received a scholarship to work with the Computer Science department at Cambridge University, which, on a side note, was something that was quite unusual for a young Egyptian and Muslim woman to do. This was in the year 2000, so it was before we all had smartphones, but at the time I was quite interested in this idea of human-computer interaction and how our interface is going to evolve over the next few years.

通过我自己的经历，我意识到我花了很多时间坐在我的机器前，在那里我编写代码并撰写所有这些研究论文，这让我意识到了两个问题。第一个问题是，我使用的笔记本电脑（记住那时还没有智能手机）应该与我非常亲密。我的意思是，我花了很多时间在它上面，虽然它知道很多关于我的事情——比如我是在写 Word 文档还是在编写代码——但它不知道我的感受。它知道我的位置，知道我的身份，但它对我的情绪和认知状态完全一无所知。

Through my own experience, I realized that I was spending a lot of time in front of my machine, where I was coding and writing all these research papers, which opened me to two realizations. The first realization was that the laptop I was using (remember no smartphones yet) was supposedly quite intimate with me. I mean, I was spending a lot of hours with it, and while it knew a lot of things about me—like if I was writing a Word document or coding—it had no idea how I was feeling. It knew my location, it knew my identity, but it was just completely oblivious to my emotional and cognitive state.

从这个意义上说，我的笔记本电脑让我想起了 Microsoft Clippy，当你正在写一篇论文时，这个回形针就会出现，稍微旋转一下，然后它会说：“哦，看起来你正在写一封信！你需要帮助吗？”Clippy 经常会在最奇怪的时间出现，例如当我压力很大，而我的截止日期是 15 分钟时......回形针就会做出它滑稽的小动作。Clippy 让我意识到我们在这里有一个机会，因为我们的技术存在情商差距。

In that sense my laptop reminded me of Microsoft Clippy, where you would be writing a paper, and then this paper-clip would show up, do a little twirl, and it would say, “Oh, it looks like you’re writing a letter! Do you need any help?” Clippy would often show up at the weirdest times, for example when I was super-stressed and my deadline was in 15 minutes... and the paperclip would do its funny little cheesy thing. Clippy helped me realize that we have an opportunity here, because there’s an emotional intelligence gap with our technology.

另一件非常清楚的事情是，这台机器帮助我与家乡的家人进行了很多交流。在攻读博士学位期间，我有时非常想家，我会和家人聊天时流着泪，但他们却不知道，因为我躲在屏幕后面。这让我感到非常孤独，我意识到，当我们进行数字互动时，我们在面对面、电话交谈或视频会议中进行的所有丰富的非语言交流都会在网络空间中消失。

The other thing that was kind of very clear is that this machine mediated a lot of my communication with my family back home. During my PhD, there were times when I was that homesick, and I would be chatting with my family in tears, and yet they’d have no idea because I was hiding behind my screen. It made me feel very lonely and I realized how all of the rich non-verbal communications that we have when we’re face to face, in a phone conversation or a video conference, are all lost in cyberspace when we are interacting digitally.

马丁·福特：所以，你自己的生活经历让你对能够理解人类情感的技术产生了兴趣。你的博士论文是否主要关注于探索这个想法？

MARTIN FORD: So, your own life experiences led you to become interested in the idea of technology that could understand human emotions. Did your PhD focus much on exploring this idea?

RANA EL KALIOUBY：是的，我对我们在技术中融入了很多智能，但情商却很少的想法很感兴趣，这是我在攻读博士学位期间开始探索的一个想法。这一切都始于我在剑桥的一次早期演讲，当时我正对观众说，我对如何制造能够读懂情绪的计算机感到好奇。在演讲中，我解释说，我自己是一个非常善于表达的人——我非常能理解人们的面部表情，而想到如何让计算机也能做到这一点，我感到非常有趣。一位博士同学突然问：“你研究过自闭症吗？自闭症患者也发现读懂面部表情和非语言行为非常困难？”由于这个问题，我在攻读博士学位期间与剑桥自闭症研究中心进行了密切合作。他们收集了令人惊叹的数据集，旨在帮助自闭症儿童了解不同的面部表情。

RANA EL KALIOUBY: Yes, I became intrigued by the idea that we’re building a lot of smartness into our technologies but not a lot of emotional intelligence, and this was an idea that I started to explore during my PhD. It all began during one of my very early presentations at Cambridge, where I was talking to an audience about how curious I was about how we might build computers that could read emotions. I explained during the presentation how I am, myself, a very expressive person—that I’m very attuned to people’s facial expressions, and how intriguing I found it to think about how we could get a computer to do the same. A fellow PhD student popped up and said, “Have you looked into autism because people on the autism spectrum also find it very challenging to read facial expressions and non-verbal behaviors?” As a result of that question, I ended up collaborating very closely with the Cambridge Autism Research Center during my PhD. They had an amazing dataset that they’d compiled to help kids on the autism spectrum to learn about different facial expressions.

机器学习需要大量数据，因此我借用了他们的数据集来训练我正在创建的算法，学习如何解读不同的情绪，这显示出了一些非常有希望的结果。这些数据为我们提供了一个机会，让我们不仅关注快乐/悲伤的情绪，还可以关注我们在日常生活中看到的许多细微的情绪，例如困惑、兴趣、焦虑或无聊。

Machine learning needs a lot of data, and so I borrowed their dataset to train the algorithms I was creating, on how to read different emotions, something that showed some really promising results. This data opened up an opportunity to focus not just on the happy/sad emotions, but also on the many nuanced emotions that we see in everyday life, such as confusion, interest, anxiety or boredom.

我很快就发现，我们有了这个工具，可以打包并作为自闭症患者的训练工具提供。这时我意识到，我的工作不仅仅是改善人机界面，还改善人际沟通和人际关系。

I could soon see that we had this tool that we could package up and provide as a training tool for individuals on the autism spectrum. This is where I realized that my work wasn’t just about improving human-computer machine interfaces, but also about improving human communication and human connection.

当我在剑桥大学完成博士学位时，我遇到了麻省理工学院的教授罗莎琳德·皮卡德 (Rosalind Picard)，她是《情感计算》一书的作者，后来她和我共同创立了 Affectiva。但早在 1998 年，罗莎琳德就认为技术需要能够识别人类的情感并对这些情感做出反应。

When I completed my PhD at Cambridge, I met with the MIT professor, Rosalind Picard, who authored the book Affective Computing, and would later co-found Affectiva with me. But back in 1998, Rosalind posited that technology needs to be able to identify human emotions and respond to those emotions.

长话短说，我们聊了起来，罗莎琳邀请我加入她在麻省理工学院媒体实验室的实验室。我之所以来到美国，是因为一个国家科学基金会的项目，该项目将利用我的阅读情绪技术，通过将其与相机相结合，我们可以将其应用于自闭症儿童。

Long story short, we ended up chatting, and Rosalind invited me to join her lab at the MIT Media Lab. The project that brought me over to the US was a National Science Foundation project that would take my technology of reading emotions and, by integrating it with a camera, we could apply it for kids on the autism spectrum.

马丁·福特：我读过一篇关于您的文章，我记得您曾描述过一种针对自闭症儿童的“情感助听器”。您指的是这个吗？这项发明还停留在概念层面，还是已经变成了一种实用产品？

MARTIN FORD: In one of the articles I read about you, I think you described an “emotional hearing aid” for autistic kids. Is this what you are referring to? Did that invention stay at the conceptual level or did it become a practical product?

RANA EL KALIOUBY：我于 2006 年加入麻省理工学院，从那时到 2009 年，我们与罗德岛州普罗维登斯的一所学校合作，他们专注于自闭症儿童。我们在那里部署了我们的技术，我们会把原型带给孩子们，让他们试用，他们会说“感觉不太对劲”，所以我们不断迭代系统，直到它开始成功。最终，我们能够证明，使用该技术的孩子有更多的眼神交流，他们做的不仅仅是看着别人的脸。

RANA EL KALIOUBY: I joined MIT in 2006, and between then and 2009 we partnered with a school in Providence, Rhode Island, and they were focused on kids on the autism spectrum. We deployed our technology there, and we would take prototypes to the kids and have them try it, and they would say “this doesn’t feel quite right,” so we iterated the system until it began to succeed. Eventually, we were able to demonstrate that the kids who were using the technology were having a lot more eye contact, and they were doing a lot more than just looking at people’s faces.

想象一下，这些患有自闭症的孩子会如何佩戴这种带有朝外摄像头的眼镜。当我们刚开始做这项研究时，我们得到的很多摄像头数据只是地板或天花板：孩子们甚至没有看脸。但是，通过与这些孩子合作，我们得到了反馈，使我们能够建立实时反馈，帮助鼓励他们进行面对面的接触。一旦这些孩子开始这样做，我们就会向他们反馈人们表现出什么样的情绪。这一切看起来都很有希望。

Imagine how these kids, somewhere on the spectrum of autism, would wear these pairs of glasses with a camera facing outwards. When we first started doing this research, a lot of the camera data we got was just of the floor or the ceiling: the kids weren’t even looking at the face. But the input that we got, from working with these kids, allowed us to build real-time feedback that helped encourage them to make face contact. Once those kids started to do that, we gave them feedback on what kind of emotions people are displaying. It all looked very promising.

你必须记住，媒体实验室是麻省理工学院的一个独特学术部门，它与行业有着非常紧密的联系，实验室 80% 的资金来自财富 500 强企业。因此，我们每年两次邀请这些公司参加所谓的“赞助商周”，这是一个“演示或死亡”的过程，因为你必须实际展示你正在做的事情。PowerPoint 根本不够用！

You’ve got to remember that Media Lab is a unique academic department at MIT, in the sense that it has very strong ties to industry, to the point where about 80% of the lab’s funding comes from Fortune 500 companies. So twice a year, we would host these companies for what we called Sponsor Week, where it was very demo-or-die because you had to actually show what you were working on. A PowerPoint wouldn’t cut it!

因此，在 2006 年至 2008 年期间，我们每年两次邀请这些人到麻省理工学院，演示自闭症原型。在这类活动中，百事可乐等公司会询问我们是否考虑过应用这项工作来测试广告是否有效。宝洁公司希望用它来测试其最新的沐浴露，因为它想知道人们是否喜欢这些气味。丰田希望用它来监测驾驶员状态，美国银行希望优化银行体验。我们考虑聘请更多研究助理来帮助开发资助者想要的想法，但我们很快意识到这不再是研究，而是一个商业机会。

So, twice a year between 2006 and 2008 we’d invite all these folks over to MIT, and we would demo the autism prototype. During these kinds of events, companies like Pepsi would ask if we’d thought about applying this work to test whether advertising was effective. And Procter & Gamble wanted to use it to test its latest shower gels, because it wanted to know if people liked the smells or not. Toyota wanted to use it for driver state monitoring, and The Bank of America wanted to optimize the banking experience. We explored getting some more research assistants to help develop the ideas that our funders wanted, but we soon realized that this was not research anymore, that it was in fact a commercial opportunity.

我对离开学术界感到不安，但我开始感到有些沮丧，因为在学术界，你可以制作所有这些原型，但它们从未大规模部署。有了公司，我觉得我们有机会扩大规模并将产品推向市场，并改变人们日常沟通和做事的方式。

I was apprehensive about leaving academia, but I was starting to get a little frustrated that in academia you do all these prototypes, but they never get deployed at scale. With a company, I felt we had an opportunity to scale and bring products to market, and to change how people communicate and do things on a day-to-day basis.

马丁·福特：听起来 Affectiva 一直非常以客户为导向。许多初创公司都试图创造一种产品，以期满足市场需求；但就您而言，客户会告诉您他们想要什么，然后您直接做出回应。

MARTIN FORD: It sounds like Affectiva has been very customer-driven. Many startups try to create a product in anticipation of a market being there; but in your case, the customers told you exactly what they wanted, and you responded directly to that.

RANA EL KALIOUBY：您说得完全正确，我们很快就意识到，我们正处于一个潜在的巨大商业机会中。总的来说，罗莎琳德和我都觉得，我们一起开创了这个领域，我们是思想领袖，而且我们也希望以一种非常合乎道德的方式来做这件事——这是我们的核心。

RANA EL KALIOUBY: You’re absolutely right, and it quickly became apparent that we were sitting on a potentially huge commercial opportunity. Collectively, Rosalind and I felt that between us we had started this field, we were thought leaders, and that we wanted to do it in a very ethical way as well—which was core to us.

马丁·福特：您现在在 Affectiva 从事什么工作？您对它未来的总体发展有什么愿景？

MARTIN FORD: What are you working on at Affectiva now, and what’s your overall vision for where it’s going to go in the future?

RANA EL KALIOUBY：我们的总体愿景是，我们肩负着让技术人性化的使命。我们开始看到技术渗透到我们生活的方方面面。我们也开始看到界面如何变得具有对话性，我们的设备如何变得更加感知化，并且更具有潜在的关联性。我们正在与我们的汽车、手机以及亚马逊的 Alexa 或苹果的 Siri 等智能设备建立紧密的关系。

RANA EL KALIOUBY: Our overall vision is that we’re on a mission to humanize technology. We’re starting to see technology permeate every aspect of our life. We’re also starting to see how interfaces are becoming conversational, and that our devices are becoming more perceptual—and a lot more potentially relational. We’re forming these tight relationships with our cars, our phones, and our smart-enabled devices like Amazon’s Alexa or Apple’s Siri.

想想那些正在制造这些设备的人，现在他们专注于这些设备的认知智能方面，而不太关注情商。但如果你看看人类，不仅仅是你的智商决定了你在职业和个人生活中的成功；它往往与你的情商和社交智力有关。你能理解周围人的心理状态吗？你能调整你的行为以考虑到这一点，然后激励他们改变他们的行为，或说服他们采取行动吗？

If you think about a lot of people who are building these devices, right now, they’re focused on the cognitive intelligence aspect of these devices, and they’re not paying much attention to the emotional intelligence. But if you look at humans, it’s not just your IQ that matters in how successful you are in your professional and personal life; it’s often really about your emotional and social intelligence. Are you able to understand the mental states of people around you? Are you able to adapt your behavior to take that into consideration and then motivate them to change their behavior, or persuade them to take action?

在我们要求人们采取行动的所有这些情况下，我们都需要有情商才能做到这一点。我认为这对于每天都会与您互动并可能要求您做事的技术来说同样如此。

All of these situations, where we are asking people to take action, we all need to be emotionally intelligent to get to that point. I think that this is equally true for technology that is going to be interfacing with you on a day-to-day basis and potentially asking you to do things.

无论是帮助您睡得更好、吃得更好、锻炼更多、工作更有效率还是更具社交能力，无论这项技术是什么，当它试图说服您参与其中时，都需要考虑您的心理状态。

Whether that is helping you sleep better, eat better, exercise more, work more productively, or be more social, whatever that technology is, it needs to consider your mental state when it tries to persuade you to take part in them.

我的论点是，这种人机界面将变得无处不在，它将根植于未来的人机界面中，无论是我们的汽车、手机还是我们家里或办公室里的智能设备。我们将与这些新设备和新类型的界面共存和协作。

My thesis is that this kind of interface between humans and machines is going to become ubiquitous, that it will just be ingrained in the future human-machine interfaces, whether it’s our car, our phone or smart devices at our home or in the office. We will just be coexisting and collaborating with these new devices, and new kinds of interfaces.

马丁·福特：你能介绍一下你正在研究的一些具体事项吗？我知道你正在做一些监控汽车司机的工作，以确保他们专心驾驶。

MARTIN FORD: Could you sketch out some of the specific things you’re working on? I know you’re doing something with monitoring drivers in cars to make sure they are attentive.

RANA EL KALIOUBY：是的，目前监控汽车驾驶员的问题在于，需要应对的情况太多了，Affectiva 公司专注于符合道德规范的情况，以及产品与市场契合度高的情况。当然，也包括市场已经做好准备的情况。

RANA EL KALIOUBY: Yes, the issue today around monitoring drivers in cars is that there are so many situations to cater for, that Affectiva as a company has focused specifically on situations that are ethical, and where there’s a good product-market fit. And of course, for where the markets are ready.

正如我之前提到的，当 Affectiva 于 2009 年成立时，第一种唾手可得的市场机会是在广告测试中，而今天 Affectiva 与四分之一的财富全球 500 强公司合作，帮助他们了解他们的广告与消费者之间建立的情感联系。

When Affectiva started in 2009, the first kind of low-hanging market opportunities were in advertising testing, as I mentioned, and today Affectiva works with a quarter of the Fortune Global 500 companies to help them understand the emotional connection their advertising creates with their consumers.

通常，公司会花费数百万美元制作一则有趣或打动人心的广告。但他们不知道这些广告是否能触动你的心弦。在我们的技术出现之前，他们唯一能找到答案的方法就是询问人们。所以，如果你，马丁·福特，是观看广告的人，那么你会收到一份调查问卷，上面会问：“嘿，你喜欢这个广告吗？你觉得它有趣吗？你会购买这个产品吗？”而问题在于，这些数据非常不可靠，而且非常有偏见。

Often, companies will spend millions of dollars to create an advertisement that’s funny or one that tugs at your heart. But they have no idea if they struck the right emotional chord with you. The only way that they could find that sort of thing out, before our technology existed, was to ask people. So, if you, Martin Ford, were the person watching the ad, then you’d get a survey, and it would say, “Hey, did you like this ad? Did you think it was funny? Are you going to buy the product?” And the problem with that is that it’s very unreliable and very biased data.

因此，现在，有了我们的技术，当您观看广告时，在征得您的同意后，它会逐一分析您的所有面部表情，并将观看同一广告的数千人的数据汇总起来。结果是一组关于人们对广告的情感反应的公正客观的数据。然后，我们可以将这些数据与客户购买意向、甚至实际销售数据和病毒式传播等因素相关联。

So now, with our technology, as you’re watching the ad, with your consent it will analyze on a moment-by-moment basis all your facial expressions and aggregate that over the thousands of people who watched that same ad. The result is an unbiased, objective set of data around how people respond emotionally to the advertising. We can then correlate that data with things like customer purchase intent, or even actual sales data and virality.

如今，我们拥有所有这些可跟踪的 KPI，并且能够将情绪反应与实际消费者行为联系起来。这是我们的产品，覆盖 87 个国家，从美国、中国到印度，也包括伊拉克和越南等较小的国家。目前，这是一款非常强大的产品，它非常出色，因为它使我们能够收集来自世界各地的数据，而且都是非常自发的数据。我认为，即使是 Facebook 和 Google 也没有这些数据，因为它不仅仅是你的个人资料照片，还包括你某天晚上坐在卧室里看洗发水广告时的情况。这就是我们拥有的数据，也是我们算法的驱动力。

Today we have all these KPIs that can be tracked, and we’re able to tie the emotional response to actual consumer behavior. That’s a product of ours that’s in 87 countries, from the US and China to India, but also smaller countries like Iraq and Vietnam. It’s a pretty robust product at this point, and it’s been amazing because it allows us to collect data from all over the world, and it’s all very spontaneous data. It’s data that, I would argue, even Facebook and Google don’t have because it’s not just your profile picture, it’s you sitting in your bedroom one night, watching a shampoo ad. That’s the data we have, and that’s what drives our algorithm.

马丁·福特：你在分析什么？主要是基于面部表情，还是也基于声音等其他因素？

MARTIN FORD: What are you analyzing? Is it mostly based on facial expressions or also on other things like voice?

RANA EL KALIOUBY：好吧，当我们刚开始的时候，我们只研究脸部，但是大约十八个月前，我们重新开始并提出一个问题：我们人类如何监测其他人类的反应？

RANA EL KALIOUBY: Well, when we first started, we worked with just the face, but about eighteen months ago we went back to the drawing board and asked: how do we as humans monitor the responses of other humans?

人们非常善于观察周围人的心理状态，我们知道我们使用的信号中大约 55% 来自面部表情和手势，而我们回应的信号中大约 38% 来自语调。因此，一个人说话的速度、音调和声音中的能量是多少。只有 7% 的信号来自文本和某人使用的实际词语选择！

People are pretty good at monitoring the mental states of the people around them, and we know that about 55% of the signals we use are in facial expression and your gestures, while about 38% of the signal we respond to is from tone of voice. So how fast someone is speaking, the pitch, and how much energy is in the voice. Only 7% of the signal is in the text and the actual choice of words that someone uses!

现在，当你想到整个情绪分析行业，这个价值数十亿美元的行业，人们收听推文、分析短信等等，它只占人类交流方式的 7%。我喜欢思考我们在这里做的事情，试图捕捉另外 93% 的非语言交流。

Now when you think of the entire industry of sentiment analysis, the multi-billion-dollar industry of people listening to tweets and analyzing text messages and all that, it only accounts for 7% of how humans communicate. What I like to think about what we’re doing here, is trying to capture the other 93% of non-verbal communication.

回到你的问题：大约十八个月前，我成立了一个语音团队，研究这些韵律副语言特征。他们会研究语调和语音事件的发生，比如你说了多少次“嗯”或者笑了多少次。所有这些语音事件都与我们所说的实际单词无关。Affectiva 技术现在将这些因素结合起来，采用我们所谓的多模态方法，将不同的模态结合起来，以真正了解一个人的认知、社交或情感状态。

So, back to your questions: about eighteen months ago I started a speech team that looks at these prosodic paralinguistic features. They would look at the tone of voice and the occurrence of speech events, such as how many times you say “um” or how many times you laughed. All of these speech events are independent of the actual words that we’re saying. Affectiva technology now combines these things and takes what we call a multimodal approach, where different modalities are combined, to truly understand a person’s cognitive, social or emotional state.

马丁·福特：你所寻找的情感指标在不同的语言和文化中是否一致，还是在不同人群中存在显著差异？

MARTIN FORD: Are the emotional indicators you look for consistent across languages and cultures, or are there significant differences between populations?

RANA EL KALIOUBY：如果你观察一个人的面部表情，甚至是语调，你会发现，其背后的表情是普遍存在的。微笑在世界各地都是微笑。然而，我们看到了一层额外的文化表达规范或规则，它们描述了人们何时表达自己的情绪，或者他们表达情绪的频率或强度。我们看到人们放大情绪、抑制情绪，甚至完全掩饰情绪的例子。我们在亚洲市场尤其能看到掩饰的迹象，例如，亚洲人不太可能表现出负面情绪。因此，在亚洲，我们看到所谓的社交微笑或礼貌微笑的出现率有所增加。这些不是喜悦的表达，而更像是在说“我承认你”，从这个意义上说，它们是一种非常社交的信号。

RANA EL KALIOUBY: If you take facial expressions or even the tone of a person’s voice, the underlying expressions are universal. A smile is a smile everywhere in the world. However, we are seeing this additional layer of cultural display norms, or rules, that depict when people portray their emotions, or how often, or how intensely they show their emotion. We see examples of people amplifying their emotions, dampening their emotions, or even masking their emotions altogether. We particularly see signs of masking in Asian markets, where Asian populations are less likely to show negative emotions, for instance. So, in Asia we see an increased incidence of what we call a social smile, or a politeness smile. Those are not expressions of joy, but are more themed around saying, “I acknowledge you,” and in that sense they are a very social signal.

总的来说，一切都是普遍的。当然，文化上也存在细微差别，而且由于我们拥有所有这些数据，我们能够建立特定地区甚至特定国家的规范。例如，我们在中国拥有如此多的数据，以至于中国有自己的规范。我们不是比较中国人对巧克力广告的反应，而是将中国人与最像他们的亚群进行比较。这种特殊方法对于我们成功监测世界各地不同文化中的情绪状态至关重要。

By and large, everything is universal. There are cultural nuances, of course, and because we have all this data, we’ve been able to build region-specific and sometimes even country-specific norms. We have so much data in China, for instance, that China is its own norm. Instead of comparing a Chinese individual’s response to say, a chocolate ad, we compare a Chinese individual to the subpopulation that’s most like them. And this particular approach has been critical to our success in monitoring emotional states in different cultures around the world.

马丁·福特：那么我猜您正在研究的其他应用是面向安全的，例如监控驾驶员或危险设备的操作人员以确保他们保持注意力？

MARTIN FORD: I guess then that other applications you’re working on are oriented toward safety, for example monitoring drivers or the operators of dangerous equipment to make sure they stay attentive?

RANA EL KALIOUBY：当然。事实上，去年我们开始收到大量来自汽车行业的关注。这真的令人兴奋，因为这是 Affectiva 的一个重大市场机会，而且我们正在为汽车行业解决两个有趣的问题。

RANA EL KALIOUBY: Absolutely. In fact in the last year we’ve started to get a ton of inbound interest from the automotive industry. It’s really exciting because it’s a major market opportunity for Affectiva and we’re solving two interesting problems for the car industry.

如今，在有主动驾驶员的汽车中，安全是一个大问题。即使我们有特斯拉这样的半自动驾驶汽车，虽然可以自动驾驶一段时间，但仍需要副驾驶来注意，安全仍将是一个问题。

In the cars of today, where there is an active driver, safety is a huge issue. And safety will continue to be an issue, even when we have semi-autonomous vehicles like Tesla that can drive themselves for a while but do still need a co-pilot to be paying attention.

使用 Affectiva 软件，我们能够监控驾驶员或副驾驶员的困倦、注意力分散、疲劳甚至醉酒等情况。如果驾驶员醉酒，我们会提醒驾驶员，甚至可能让汽车进行干预。干预措施可以是任何方式，从改变音乐到吹出一点冷风，或系紧安全带，甚至可能说：“你知道吗？我就是汽车，我觉得我现在可以比你更安全。我来控制它。”一旦汽车了解了驾驶员的注意力水平和醉酒程度，它就可以采取很多行动。所以，这是一类用例。

Using Affectiva software, we’re able to monitor the driver or the co-pilot for things like drowsiness, distraction, fatigue and even intoxication. In the case of intoxication, we would alert the driver or also even potentially have the car intervene. Intervention could be anything from changing the music to blasting a little bit of cold air, or tightening the seat belt, all the way to potentially saying, “You know what? I’m the car, and I feel I could be a safer driver than you are right now. I’m taking control over.” There’s a lot of actions the car can take once it understands the level of attention and how impaired a driver is. So, that’s one class of use cases.

我们正在为汽车行业解决的另一个问题是乘客体验。让我们展望一下未来，那时我们将拥有完全自动驾驶汽车和机器人出租车，车上根本没有司机。在这些情况下，汽车需要了解乘客的状态，例如，车上有多少人，他们的关系如何，他们是否在交谈，甚至车上是否有可能被遗忘的婴儿？一旦您了解了车内乘客的情绪，您就可以个性化体验。

The other problem we’re solving for the automotive industry is around the occupant experience. Let’s look into the future where we have fully autonomous vehicles and robot-taxis, where there’s no driver in the car at all. In those situations, the car needs to understand the state of the occupants such as, how many people are in the car, what’s their relationship, are they in a conversation, or even do we have a baby in the car that’s potentially getting left behind? Once you understand the mood of the occupants in the car, you can personalize the experience.

机器人出租车可以推荐产品或路线。这还将为汽车公司引入新的商业模式，尤其是宝马或保时捷等高端品牌，因为目前它们只注重驾驶体验。但未来，这将不再是关于驾驶：而是关于改变和重新定义交通方式、移动体验。现代交通是一个非常令人兴奋的市场，我们投入了大量精力为该行业以及与一线公司合作的公司打造产品。

The robot-taxi could make product recommendations or route recommendations. This would also introduce new business models for auto companies, especially premium brands like a BMW or a Porsche, because right now they’re all about the driving experience. But in the future, it’s not going to be about driving anymore: it’s going to be about transforming and redefining that transport, that mobility experience. Modern transport is a very exciting market, and we’re spending a lot of our mindshare building products for that industry, and also for those partnered with Tier 1 companies.

马丁·福特：您认为它在医疗保健领域有应用潜力吗？考虑到我们确实面临心理健康危机，我想知道您是否认为您在 Affectiva 开发的技术可能在咨询等领域有所帮助？

MARTIN FORD: Do you see potential applications in healthcare? Given that we do have a mental health crisis, I wonder if you think the kind of technology you’re building at Affectiva might help in areas like counseling?

RANA EL KALIOUBY：医疗保健可能是我最感兴趣的领域，因为我们知道抑郁症有面部和声音生物标记，而且我们知道有些迹象可以预测一个人的自杀意图。想想我们坐在设备和手机前的频率，这是一个收集非常客观数据的机会。

RANA EL KALIOUBY: Healthcare is probably what I’m most excited about, because we know that there are facial and vocal biomarkers of depression, and we know that there are signs that could be predictive of suicidal intent in a person. Think about how often we are in front of our devices and our phones, that’s an opportunity to collect very objective data.

目前，你只能询问一个人，从 1 到 10 的量表，他们的抑郁程度或自杀倾向有多大。这很不准确。但我们现在有机会大规模收集数据，并建立一个基线模型，了解一个人是谁，以及他们的基本精神状态或心理健康状态是什么。一旦我们有了这些数据，如果某人开始偏离正常基线，那么系统就可以向这个人本人、他的家人甚至可能是医疗专业人员发出信号。

Right now, you can only ask a person, on a scale from 1 to 10, how depressed they are, or how suicidal they are. It’s just not accurate. But we now have the opportunity to collect data at scale and build a baseline model of who someone is and what their baseline mental state or mental health state is. Once we have that data, if someone starts to deviate from their normal baseline, then a system can signal that to the person themselves, to their family members or even maybe a healthcare professional.

然后想象一下，我们可以如何使用相同的指标来分析不同治疗方法的疗效。患者可以尝试认知行为疗法或某些药物，随着时间的推移，我们将能够非常准确、非常客观地量化这些治疗是否有效。我觉得这确实有可能理解焦虑、压力和抑郁，并能够对其进行量化。

Then imagine how we could use these same metrics to analyze the efficacy of different treatments. The person could try cognitive behavioral therapy or certain drugs, and we would be able to quantify, very accurately and very objectively over time, if those treatments were effective or not. I feel that there’s a real potential here to understand anxiety, stress, and depression, and be able to quantify it.

马丁·福特：我想谈谈人工智能的伦理问题。人们很容易想到这种技术令人不安的地方。例如，在谈判期间，如果你的系统秘密监视某人并向对方提供有关其反应的信息，这将产生不公平的优势。或者它可以用于某种形式的更广泛的工作场所监控。在某人开车时监视他们以确保他们专心，大多数人可能都可以接受，但对于你的系统监视坐在电脑前的办公室工作人员的想法，他们可能会有截然不同的看法。你如何解决这些担忧？

MARTIN FORD: I want to move into a discussion about the ethics of AI. It’s easy to think of things that people might find disturbing about this kind of technology. For example, during a negotiation, if your system was secretly watching someone and giving the other side information about their responses, that would create an unfair advantage. Or it could be used for some form of wider workplace surveillance. Monitoring someone when they’re driving to make sure they’re attentive would probably be okay with most people, but they might feel very different about the idea of your system watching an office worker sitting in front of a computer. How do you address those concerns?

RANA EL KALIOUBY：这里有一点历史教训，当 Rosalind、我和我们的第一位员工在 Rosalind 的厨房餐桌旁见面时，我们在想：Affectiva 将要接受测试，那么我们的界限是什么，什么是不可协商的？最后，我们确定了这一核心价值观，即尊重人们的情绪是一种非常私人的数据。从那时起，我们同意我们只会接受人们明确同意并选择分享这些数据的情况。而且，理想情况下，他们也会因分享这些数据而获得一些回报。

RANA EL KALIOUBY: There’s a little history lesson here about when Rosalind, myself, and our first employee met around Rosalind’s kitchen table and we were thinking: Affectiva is going to get tested, so what are our boundaries and what’s non-negotiable? In the end, we landed on this core value of respecting that people’s emotions are a very personal type of data. From then on, we agreed that we would only take on situations where people are explicitly consenting and opting in to share that data. And, ideally, where they’re also getting some value in return for sharing that data.

Affectiva 已经对这些技术进行了测试。2011 年，我们资金短缺，但我们有机会从一家拥有风险投资部门的安全机构获得资金，该机构对使用该技术进行监控和安全非常感兴趣。尽管大多数人都知道，当他们去机场时，他们会受到监视，但我们只是觉得这不符合我们的核心价值观——同意和选择加入，所以尽管有资金，我们还是拒绝了该提议。在 Affectiva，我们一直远离那些我们认为人们不一定选择加入且价值方程不平衡的应用程序。

These are things that Affectiva has been tested on. In 2011, we were running low on funds, but we had the opportunity for funding from a security agency that had a venture arm, and it was very interested in using the technology for surveillance and security. Even though most people know that when they go to an airport, they’re being watched, we just felt that this was not in line with our core value of consent and opt-in, so we declined the offer even though the money was there. At Affectiva, we’ve stayed away from applications where we feel that people aren’t necessarily opting in and the value equation is not balanced.

当你考虑工作场所周围的应用时，这个问题确实变得非常有趣，因为同样的工具可以以非常赋予权力的方式使用——或者当然，非常像老大哥。我确实认为，如果人们愿意匿名选择加入，雇主能够获得情绪评分，或者只是总体了解人们在办公室是否感到压力，或者人们是否投入和快乐，那将非常有趣。

When you think about the applications around the workplace, this question does become very interesting because the same tool could be used in ways that might be very empowering—or of course, very like Big Brother. I do think it would be super-interesting if people wanted to opt-in, anonymously, and employers were able to then get a sentiment score, or just an overall view, of whether people are stressed in the office—or whether people are engaged and happy.

另一个很好的例子是，一位 CEO 正在向来自世界各地的人们做演讲，机器会显示演讲信息是否符合 CEO 的意图。目标是否令人兴奋？员工是否受到激励？这些都是核心问题，如果我们都在同一地点工作，就很容易收集这些问题；但现在，由于每个人都分散在各地，很难了解这些问题。然而，如果你反过来用同样的技术说：“好吧。我要挑某位员工说点什么，因为他们似乎真的很不投入”，那么这就是对数据的完全滥用。

Another great example would be where a CEO is giving a presentation, to people dialed in from around the world, and the machine indicates whether or not the message is resonating as they CEO intends. Are the goals exciting? Are people motivated? These are core questions that if we’re all co-located, it would be easy to collect; but now, with everybody distributed, it’s just really hard to get a sense of these things. However, if you turn it around and use the same technology to say, “OK. I’m going to pick on a certain member of staff because they seemed really disengaged,” then that’s a total abuse of the data.

另一个例子是，我们有一种技术可以跟踪会议进展情况，每次会议结束时，它都可以给人们反馈。它会给你这样的反馈：“你喋喋不休地说了 30 分钟，而且你对某某非常有敌意，你应该多一点体贴或多一点同理心。”你可以很容易地想象，这项技术可以作为教练，帮助员工更好地谈判或成为更有思想的团队成员；但与此同时，你也可以用它来损害人们的职业生涯。

Another example would be where we have a version of the technology that tracks how meetings go, and at the end of every meeting, it can give people feedback. It would give you feedback like, “you rambled for 30 minutes, and you were pretty hostile towards so-and-so, you should be a little bit more thoughtful or more empathetic.” You can easily imagine how this technology could be used as a coach to help staff negotiate better or be a more thoughtful team member; but at the same time, you could use it to hurt people’s careers.

我想，我们应该提倡这样一种情况，人们可以取回数据，从中学到一些东西，这可以帮助他们提高社交和情商技能。

I would like to think of us as advocating for situations where people can get the data back, and they can learn something about it, and it could help them advance their social and emotional intelligence skills.

马丁·福特：让我们深入探讨一下您使用的技术。我知道您大量使用深度学习。您如何看待这项技术？最近出现了一些反对意见，一些人认为深度学习的进展将会放缓甚至陷入困境，需要采用另一种方法。您如何看待神经网络的使用以及它们在未来将如何发展？

MARTIN FORD: Let’s delve into the technology you’re using. I know that you use deep learning quite heavily. How do you feel about that as a technology? There has been some recent pushback, with some people suggesting that progress in deep learning is going to slow or even hit a wall, and that another approach will be needed. How do you feel about the use of neural networks and how they’re going to evolve in the future?

RANA EL KALIOUBY：当我读博士的时候，我使用动态贝叶斯网络来量化和构建这些分类器。几年前，我们将所有科学基础设施都转移到深度学习上，我们确实从中获益匪浅。

RANA EL KALIOUBY: Back when I did my PhD, I used dynamic Bayesian networks to quantify and build these classifiers. Then a couple of years ago we moved all our science infrastructure to be deep learning-based, and we have absolutely reaped the benefits of that.

我想说，深度学习还远未达到极限。随着更多数据与这些深度神经网络的结合，我们在很多不同情况下的分析准确性和稳健性都有所提高。

I would say that we haven’t even maxed out yet on deep learning. With more data combined with these deep neural nets, we see increases in the accuracy and robustness of our analysis across so many different situations.

深度学习很棒，但我不认为它能满足我们所有的需求。它仍然在很大程度上是监督式的，所以你仍然需要一些标记数据来跟踪这些分类器。我认为它是机器学习这个大领域中很棒的工具，但深度学习不会是我们使用的唯一工具。

Deep learning being awesome, I don’t think that it’s the be-all, end-all to all of our needs. It’s still pretty much supervised, so you still need to have some labeled data to track these classifiers. I think of it as an awesome tool within this bigger bucket of machine learning, but deep learning is not going to be the only tool that we use.

马丁·福特：现在让我们从更宏观的角度来谈谈通用人工智能的发展。其中的障碍是什么？通用人工智能是可行的、现实的，还是你希望在有生之年看到的东西？

MARTIN FORD: Thinking more generally now, let’s talk about the march towards artificial general intelligence. What are the hurdles involved? Is AGI something that is feasible, realistic or even something you expect to see in your lifetime?

RANA EL KALIOUBY：我们距离 AGI 还有很多年，我之所以这么说，是因为当你看到我们今天拥有的所有 AI 示例时，你会发现它们都非常狭窄。今天的 AI 擅长做一件事，但它们都必须以某种方式引导，即使它们学会了如何从头开始玩游戏。

RANA EL KALIOUBY: We’re many, many, many, many, many years away from an AGI and the reason I say that is because when you look at all the examples of AI that we have today, all of them are pretty narrow. Today’s AI does one thing well, but they all had to be bootstrapped in one way or another, even if they learned how to play a game from scratch.

我认为数据集中存在一些子假设或某种程度的子管理，这使得算法可以学习它所学到的任何内容，但我认为我们还没有弄清楚如何赋予它人类水平的智能。

I think there are sub-assumptions, or some level of sub-curation, that happens with the dataset, which has allowed that algorithm to learn whatever it learns, and I don’t think that we’ve yet figured out how to give it human-level intelligence.

即使你使用当今最好的自然语言处理系统，并且给它进行类似三年级测试的测试，它也不会通过。

Even if you look at the best natural language processing system that we have today, and you give it something like a third-grade test, it doesn’t pass.

马丁·福特：您如何看待 AGI 和情感之间的交集？您的许多工作主要集中在让机器理解情感，但反过来想，拥有一台能够表现出情感的机器又如何呢？您认为这是 AGI 的重要组成部分吗？或者您想象中的机器就像僵尸一样，完全没有情感意识？

MARTIN FORD: What are your thoughts about the intersection between AGI and emotion? A lot of your work is primarily focused on getting machines to understand emotion, but flipping the coin, what about having a machine that exhibits emotion? Do you think that’s an important part of what AGI would be, or do you imagine a zombie-like machine that has no emotional sense at all?

RANA EL KALIOUBY：我认为，就机器表现出情感而言，我们现在已经做到了这一点。Affectiva 开发了一个情感感知平台，我们的许多合作伙伴都使用这个感知平台来驱动机器行为。无论这项技术是汽车还是社交机器人，情感感知平台都可以将我们的人类指标作为输入，这些数据可用于决定机器人如何响应。这些响应可能是机器人根据我们的刺激说出的话，就像亚马逊 Alexa 今天做出的响应一样。

RANA EL KALIOUBY: I would say that we are already there, right now, in terms of machines exhibiting emotions. Affectiva has developed an emotion-sensing platform, and a lot of our partners use this sensing platform to actuate machine behavior. Whether that technology is a car, or a social robot, an emotion-sensing platform can take our human metrics as input, and that data can be used to decide how a robot is going to respond. Those responses could be the things that a robot says from our stimuli, just like Amazon Alexa responds today.

当然，如果你让亚马逊 Alexa 订购某样东西，但它总是出错，那么你现在就会很恼火。但 Alexa 不会对此一无所知，你的 Alexa 设备可以说：“好的，我很抱歉。我意识到我错了。让我再试一次。” Alexa 可以了解我们的沮丧程度，然后可以将其纳入其响应中，并融入到它接下来的实际行动中。机器人可以移动头部，可以四处走动，可以书写，还可以表现出我们可以理解的动作，“哦！它看起来很抱歉。”

Of course, if you’re asking Amazon Alexa to order something and it keeps getting it wrong, then you’re now getting annoyed. But instead of Alexa just being completely oblivious to all of that, your Alexa device could say, “OK, I’m sorry. I realize I’m getting this wrong. Let me try again.” Alexa could acknowledge our level of frustration and it could then incorporate that into its response, and into what it actually does next. A robot could move its head, it could move around, it could write, and it could exhibit actions that we would translate into, “Oh! It looks like it’s sorry.”

我认为机器系统已经将情感暗示融入到其行为中，并且它们能够以任何设计的方式表现出情感。当然，这与设备真正拥有情感完全不同，但我们不需要这样做。

I would argue that machine systems are already incorporating emotional cues in their actions, and that they can portray emotions, in any way that someone designs them to do so. That is quite different, of course, from the device actually having emotions, but we don’t need to go there.

马丁·福特：我想谈谈对就业的潜在影响。您对此有何看法？您是否认为人工智能和机器人技术有可能对经济和就业市场造成巨大冲击，或者您是否认为这可能被夸大了，我们不应该太过担心？

MARTIN FORD: I want to talk about the potential impact on jobs. How do you feel about that? Do you think that there is the potential for a big economic and job-market disruption from AI and robotics, or do you think that’s perhaps been overhyped, and we shouldn’t worry quite so much about it?

RANA EL KALIOUBY：我更愿意将其视为人类与技术的合作。我承认有些工作将不复存在，但这在人类历史上并不是什么新鲜事。我们已经一次又一次地见证了这种工作转移，因此我认为将会出现一类全新的工作和就业机会。虽然我们现在可以设想一些新的工作，但我们无法设想所有新的工作。

RANA EL KALIOUBY: I’d like to think of this as more of a human-technology partnership. I acknowledge that some jobs are going to cease to exist, but that’s nothing new in the history of humanity. We’ve seen that shift of jobs over and over again, and so I think there’s going to be a whole new class of jobs and job opportunities. While we can envision some of those new jobs now, we can’t envision all of them.

我不认同这样的世界观：机器人​​将接管一切，控制一切，而人类则只是坐在海边闲坐。我在第一次海湾战争期间在中东长大，所以我意识到世界上有太多问题需要解决。我不认为我们离一台某天醒来就能解决所有这些问题的机器还很远。所以，回答你的问题，我并不担心。

I don’t subscribe to the vision of a world where robots are going to take over and be in control, whilst humanity will just sit around and chill by the beach. I grew up in the Middle East during the time of the first Gulf War, so I’ve realized that there are so many problems in the world that need to be solved. I don’t think we’re anywhere close to a machine that’s just going to wake up someday and be able to solve all these problems. So, to answer your question, I’m not concerned.

马丁·福特：如果你考虑一下相对常规的工作，例如呼叫中心的客户服务工作，那么你正在开发的技术似乎也能够让机器完成更人性化的工作。当我被问到这个问题时，我经常会说，最有可能安全的工作是更以人为本的工作，那些涉及情商的工作。但听起来你也在将技术推向这个领域，所以看起来有很多职业最终都可能受到影响，包括一些目前被认为相当安全的领域。

MARTIN FORD: If you think about a relatively routine job, for example a customer service job in a call center, it does sound like the technology you’re creating might enable machines to do that more human element of the work as well. When I’m asked about this, which is often, I say the jobs that are most likely to be safe are the more human-oriented jobs, the ones that involve emotional intelligence. But it sounds like you’re pushing the technology into this area as well, so it does seem that there’s a very broad range of occupations that could be eventually be impacted, including some areas currently perceived as quite safe from automation.

RANA EL KALIOUBY：我认为你说得对，让我举个护士的例子。在 Affectiva，我们正在与一些公司合作，为我们的手机打造护士化身，甚至在我们的家中安装社交机器人，这些机器人旨在陪伴绝症患者。我不认为这会取代真正的护士，但我确实认为它会改变护士的工作方式。

RANA EL KALIOUBY: I think you’re right about this, and let me give an example with nurses. At Affectiva, we are collaborating with companies that are building nurse avatars for our phones, and even installing social robots in our homes, which are designed to be a companion to terminally-ill patients. I don’t think this is going to take the place of real nurses, but I do think it’s going to change how nurses do their jobs.

你可以很容易地想象，一名人类护士可以负责 20 名患者，而每名患者都可以使用护士化身或护士机器人。只有当护士机器人无法处理问题时，人类护士才会介入。这项技术使护士机器人能够管理更多患者，并以一种目前不可能的方式进行纵向管理。

You can easily imagine how a human nurse could be assigned to twenty patients, and each of these patients has access to a nurse avatar or a nurse robot. The human nurse only gets brought into the loop if there is a problem that the nurse robot can’t deal with. The technology allows the nurse robot to manage so many more patients, and manage them longitudinally, in a way that’s not possible today.

教师也有类似的例子。我不认为智能学习系统会取代教师，但它们会在缺乏足够教师的地方增强教师的作用。这就像我们将这些工作委托给那些可以代替我们完成部分工作的微型机器人。

There’s a similar example with teachers. I don’t think intelligent learning systems are going to replace teachers, but they are going to augment them in places where there isn’t access to enough teachers. It’s like we’re delegating these jobs to those mini-robots that could do parts of the job on our behalf.

我认为卡车司机也是如此。未来十年内，没有人会开卡车，但有人会坐在家里远程操控 100 个车队，确保它们都在正轨上。相反，可能会有这样一种工作，需要有人时不时地介入，对其中一辆卡车进行人工控制。

I think this is even true for truck drivers. Nobody will be driving a truck in the next ten years, but someone is sitting at home and tele-operating 100 fleets out there and making sure that they’re all on track. There may instead be a job where someone needs to intervene, every so often, and take human control of one of them.

马丁·福特：您对人们对人工智能或通用人工智能所表达的一些担忧有何回应，尤其是伊隆·马斯克，他曾直言不讳地谈论过生存风险？

MARTIN FORD: What is your response to some of the fears expressed about AI or AGI, in particular by Elon Musk, who has been very vocal about existential risks?

RANA EL KALIOUBY：互联网上有一部纪录片，名为《你相信这台电脑吗？》，该片部分资金由伊隆·马斯克提供，我在片中接受了采访。

RANA EL KALIOUBY: There’s a documentary on the internet called Do You Trust This Computer? which was partially funded by Elon Musk, and I was featured in it being interviewed.

马丁·福特：是的，事实上，我在这本书中采访过的其他几个人也出现在那部纪录片中。

MARTIN FORD: Yes, in fact, a couple of the other people I’ve interviewed in this book were also featured in that documentary.

RANA EL KALIOUBY：我在中东长大，我觉得人类面临的问题比人工智能更大，所以我并不担心。

RANA EL KALIOUBY: Having grown up in the Middle East, I feel that humanity has bigger problems than AI, so I’m not concerned.

我认为，这种认为机器人将接管人类的生存威胁的观点剥夺了我们作为人类的自主权。归根结底，我们正在设计这些系统，我们可以决定如何部署它们，我们可以关闭开关。所以，我不认同这些恐惧。我确实认为我们对人工智能有更迫切的担忧，这些担忧与人工智能系统本身有关，以及我们是否通过它们延续了偏见？

I feel that this view, about the existential threat that robots are going to take over humanity, takes away our agency as humans. At the end of the day, we’re designing these systems, and we get to say how they are deployed, we can turn the switch off. So, I don’t subscribe to those fears. I do think that we have more imminent concerns with AI, and these have to do with the AI systems themselves and whether we are, through them, just perpetuating bias?

马丁·福特：那么，您会说偏见是我们目前面临的最紧迫的问题之一吗？

MARTIN FORD: So, you would say that bias is one of the more pressing issues that we’re currently facing?

RANA EL KALIOUBY：是的。由于技术发展如此之快，我们在训练这些算法时，并不一定确切知道算法或神经网络正在学习什么。我担心，通过将这些偏见应用到这些算法中，我们只是在重建社会中存在的所有偏见。

RANA EL KALIOUBY: Yes. Because the technology is moving so fast, while we train these algorithms, we don’t necessarily know exactly what the algorithm or the neural network is learning. I fear that we are just rebuilding all the biases that exist in society by implementing them in these algorithms.

马丁·福特：因为数据来自人类，所以不可避免地会包含人类的偏见。你说有偏见的不是算法，而是数据。

MARTIN FORD: Because the data is coming from people, so inevitably it incorporates their biases. You’re saying that it isn’t the algorithms that are biased, it’s the data.

RANA EL KALIOUBY：没错，关键在于数据。关键在于我们如何应用这些数据。因此，Affectiva 作为一家公司，非常透明地表明我们需要确保训练数据代表所有不同种族群体，并且具有性别平衡和年龄平衡。

RANA EL KALIOUBY: Exactly, it’s the data. It’s how we’re applying this data. So Affectiva, as a company, is very transparent about the fact that we need to make sure that the training data is representative of all the different ethnic groups, and that it has gender balance and age balance.

我们需要非常认真地考虑如何训练和验证这些算法。这是一个持续关注的问题，它始终是一项正在进行的工作。我们总能做更多的事情来防范这些偏见。

We need to be very thoughtful about how we train and validate these algorithms. This an ongoing concern, it’s always a work in progress. There is always more that we can do to guard against these kinds of biases.

马丁·福特：但积极的一面是，虽然纠正人的偏见非常困难，但一旦你理解了算法的偏见，纠正算法的偏见可能会容易得多。你可以很容易地提出这样的论点：未来更多地依赖算法可能会让世界减少偏见和歧视。

MARTIN FORD: But the positive side would be that while fixing bias in people is very hard, fixing bias in an algorithm, once you understand it, might be a lot easier. You could easily make an argument that relying on algorithms more in the future might lead to a world with much less bias or discrimination.

RANA EL KALIOUBY：没错。一个很好的例子就是招聘。Affectiva 与一家名为 HireVue 的公司合作，该公司在招聘过程中使用我们的技术。求职者无需发送 Word 简历，而是发送视频面试，通过结合我们的算法和自然语言处理分类器，系统会根据求职者的非语言交流以及他们回答问题的方式对他们进行排名和排序。该算法不考虑性别，也不考虑种族。因此，这些面试的第一个筛选条件不会考虑性别和种族。

RANA EL KALIOUBY: Exactly. One great example is in hiring. Affectiva has partnered a company called HireVue, who use our technology in the hiring process. Instead of sending a Word resume, candidates send a video interview, and by using a combination of our algorithms and natural language processing classifiers, the system ranks and sorts those candidates based on their non-verbal communication, in addition to how they answered the questions. This algorithm is gender-blind, and it’s ethnically blind. So, the first filters for these interviews do not consider gender and ethnicity.

HireVue 发布了与联合利华合作的案例研究，其中表明，联合利华不仅将招聘时间缩短了 90%，而且招聘流程还使其新招聘人员的多样性提高了 16%。我觉得这很酷。

HireVue has published a case study, with Unilever, where it shows that not only did it reduce its time to hire by 90%, but the process resulted in a 16% increase in the diversity of its incoming hiring population. I found that to be pretty cool.

马丁·福特：您认为人工智能需要监管吗？您曾谈到 Affectiva 的道德标准非常高，但展望未来，您的竞争对手很有可能会开发类似的技术，但可能不会遵守相同的标准。他们可能会接受专制国家的合同，或者接受想要秘密监视其员工或客户的公司的合同，即使您不会这样做。鉴于此，您认为有必要对这类技术进行监管吗？

MARTIN FORD: Do you think AI will need to be regulated? You’ve talked about how you’ve got very high ethical standards at Affectiva, but looking into the future, there’s a real chance that your competitors are going to develop similar technologies but perhaps not adhere to the same standards. They might accept the contract from an authoritarian state, or the corporation that wants to secretly spy on its employees or customers, even if you would not. Given this, do you think there’s going to be a need to regulate this type of technology?

RANA EL KALIOUBY：我非常支持监管。Affectiva 是人工智能联盟合作伙伴的一部分，也是 FATE 工作组的成员，该工作组是公平、负责、透明和平等的人工智能组织。

RANA EL KALIOUBY: I’m a big advocate of regulation. Affectiva is part of the Partnership on AI consortium, and a member of the FATE working group, which is the Fair, Accountable, Transparent and Equitable AI.

通过与这些团体合作，我们的任务是制定指导方针，倡导与 FDA（食品和药物管理局）同等的 AI 流程。除了这项工作之外，Affectiva 还发布了行业最佳实践和指导方针。由于我们是思想领袖，我们有责任成为监管倡导者，推动事情向前发展，而不是只是说：“哦，是的。我们只是要等到立法出台。”我不认为这是正确的解决方案。

Through working with these groups, our mandate is to develop guidelines that advocate for the equivalent of an FDA (Food and Drug Administration) process for AI. Alongside this work, Affectiva publishes best practices and guidelines for the industry. Since we are thought leaders, it is our responsibility to be an advocate for regulation, and to move the ball forward, as opposed to just saying, “Oh, yeah. We’re just going to wait until legislation comes about.” I don’t think that that’s the right solution.

我也是世界经济论坛的成员，论坛上有一个关于机器人和人工智能的国际论坛理事会。通过与这个论坛合作，我对不同国家对人工智能的看法的文化差异产生了浓厚的兴趣。中国就是一个典型的例子，中国是该理事会的成员。我们知道中国政府并不真正关心道德问题，因此这就引出了一个问题：你如何应对这种情况？不同的国家对人工智能监管的看法不同，这使得这个问题很难回答。

I’m also a part of the World Economic Forum, on which there’s an international forum council on robotics and AI. Through working with this forum, I’ve become fascinated by the cultural differences in how different countries think about AI. A great example can be seen in China, which is part of this council. We know that the Chinese government doesn’t really care about ethics, and so it begs the question, how do you navigate that? Different nations think about AI regulation differently, which makes this difficult to answer the question.

马丁·福特：最后，我想您是个乐观主义者吧？您相信这些技术总体上将对人类有益吗？

MARTIN FORD: To end on an upbeat note, I assume you’re an optimist? That you believe these technologies are, on balance, going to be beneficial for humanity?

RANA EL KALIOUBY：是的，我会说我是一个乐观主义者，因为我相信技术是中性的。重要的是我们决定如何使用它，我认为它有好的潜力，作为一个行业，我们应该追随我团队的脚步，我们决定将我们的注意力集中在人工智能的积极应用上。

RANA EL KALIOUBY: Yes, I would say that I’m an optimist because I believe that technology is neutral. What matters is how we decide to use it, and I think there’s a potential for good, and we should, as an industry, follow the footsteps of my team, where we’ve decided to focus our mindshare on the positive applications of AI.

RANA EL KALIOUBY 是专注于情感 AI 的公司 Affectiva 的首席执行官兼联合创始人。她在埃及开罗的美国大学获得了本科和硕士学位，并在剑桥大学计算机实验室获得了博士学位。她曾在麻省理工学院媒体实验室担任研究科学家，开发了帮助自闭症儿童的技术。这项工作直接促成了 Affectiva 的成立。

RANA EL KALIOUBY is the CEO and co-founder of Affectiva, a company focused on emotion AI. She received her undergraduate and master’s degrees from American University in Cairo, Egypt and her PhD from the Computer Lab at the University of Cambridge. She worked as a research scientist at the MIT Media Lab, where she developed technology to assist autistic children. That work led directly to the launch of Affectiva.

Rana 获得过许多奖项和荣誉，包括 2017 年被世界经济论坛评选为全球青年领袖。她还入选了《财富》杂志的“ 40 位 40 岁以下精英”榜单 和 TechCrunch 的 “2016 年 40 位杰出女性创始人” 榜单。

Rana has received a number of awards and distinctions, including selection as a Young Global Leader in 2017 by the World Economic Forum. She was also featured on Fortune Magazine’s 40 under 40 and TechCrunch’s 40 Female founders who crushed it in 2016 lists.

谷歌工程总监

DIRECTOR OF ENGINEERING AT GOOGLE

雷·库兹韦尔是世界领先的发明家、思想家和未来学家之一。他曾获得 21 个荣誉博士学位，并被三任美国总统授予荣誉称号。他是麻省理工学院莱默森创新奖的获得者，并于 1999 年获得克林顿总统颁发的国家技术奖章，这是美国技术领域的最高荣誉。雷也是一位多产的作家，创作了 5 本全国畅销书。2012 年，雷成为谷歌的工程总监，领导一支工程师团队开发机器智能和自然语言理解。雷的第一部小说《丹尼尔，女超人编年史》将于 2019 年初出版。雷的另一本书《奇点越来越近》预计将于 2019 年底出版。

Ray Kurzweil is one of the world’s leading inventors, thinkers, and futurists. He has received 21 honorary doctorates, and honors from three US presidents. He is the recipient of the MIT Lemelson Prize for innovation and in 1999, he received the National Medal of Technology, the nation’s highest honor in technology, from President Clinton. Ray is also a prolific writer, authoring 5 national bestsellers. In 2012, Ray became a Director of Engineering at Google—heading up a team of engineers developing machine intelligence and natural language understanding. Ray’s first novel, Danielle, Chronicles of a Superheroine, is being published in early 2019. Another book by Ray, The Singularity is Nearer, is expected to be published in late 2019.

马丁·福特：您是如何开始涉足人工智能领域的？

MARTIN FORD: How did you come to start out in AI?

雷·库兹韦尔：我第一次接触人工智能是在 1962 年，这距离马文·明斯基和约翰·麦卡锡在 1956 年新罕布什尔州汉诺威举行的达特茅斯会议上提出这一术语仅有 6 年。

RAY KURZWEIL: I first got involved in AI in 1962, which was only 6 years after the term was coined by Marvin Minsky and John McCarthy at the 1956 Dartmouth Conference in Hanover, New Hampshire.

人工智能领域已经分裂为两个互相争斗的阵营：符号学派和联结学派。符号学派绝对占据优势地位，马文·明斯基被视为其领袖。联结学派是新贵，其中一位是康奈尔大学的弗兰克·罗森布拉特，他发明了第一个流行的神经网络，即感知器。我给他们两人都写了信，他们都邀请我去，所以我先去拜访明斯基，他和我呆了一整天，我们建立了持续 55 年的融洽关系。我们谈到了人工智能，当时这是一个非常模糊的领域，没有人真正关注它。他问我接下来要见谁，当我提到罗森布拉特博士时，他说我不必费心了。

The field of AI had already bifurcated into two warring camps: the symbolic school and the connectionist school. The symbolic school was definitely in the ascendancy with Marvin Minsky regarded as its leader. The connectionists were the upstarts, and one such person was Frank Rosenblatt at Cornell University, who had the first popularized neural net called the perceptron. I wrote them both letters and they both invited me to come up, so I first went to visit Minsky, where he spent all day with me and we struck up a rapport that would last for 55 years. We talked about AI, which at the time was a very obscure field that nobody was really paying attention to. He asked who I was going to see next, and when I mentioned Dr. Rosenblatt, he said that I shouldn’t bother.

然后我去见了罗森布拉特博士，他有一个单层神经网络，称为感知器；这是一个带有摄像头的硬件设备。我带了一些打印的信件去见罗森布拉特博士，只要这些信件是 Courier 10 格式的，他的设备就能完美识别它们。

I then went to go and see Dr. Rosenblatt, who had this single-layer neural net called the perceptron; it was a hardware device that had a camera. I brought some printed letters to my meeting with Dr. Rosenblatt where his device recognized them perfectly as long as they were in Courier 10.

其他类型的样式效果不太好，他说：“别担心，我可以把感知器的输出作为输入提供给次级感知器，然后我们可以把其输出提供给第三层，随着我们添加层数，它会变得更聪明，更具有概括性，能够做所有这些了不起的事情。”我回答说：“你试过吗？”他说：“嗯，还没有，但这是我们研究议程上的重中之重。”

Other type styles didn’t work as well, and he said, “Don’t worry, I can take the output of the perceptron and feed it as the input to a secondary perceptron, then we can take the output of that and feed it to a third layer, and as we add layers it’ll get smarter and generalize and be able to do all these remarkable things.” I responded saying, “Have you tried that?”, and he said, “well, not yet, but it’s high on our research agenda.”

20 世纪 60 年代，事情的发展速度不如今天这么快，遗憾的是，他于 9 年后（1971 年）去世，从未尝试过这个想法。然而，这个想法非常有先见之明。我们现在在神经网络中看到的所有激动人心的事情都归功于这些具有多层的深度神经网络。这是一个非常了不起的见解，因为它确实没有明显的作用。

Things didn’t move quite as quickly back in the 1960s as they do today, and sadly he died 9 years later in 1971 never having tried that idea. The idea was remarkably prescient, however. All of the excitement we see now in neural nets is due to these deep neural networks with many layers. It was a pretty remarkable insight, as it really was not obvious that it would work.

1969 年，明斯基与同事西摩·派普特 (Seymour Papert) 共同撰写了《感知器》一书。该书基本上证明了一个定理，即感知器无法设计出需要使用 XOR 逻辑函数的答案，也无法解决连通性问题。该书的封面上有两张迷宫状的图像，如果仔细观察，您会发现一张是完全连通的，另一张则不是。进行这种分类称为连通性问题。该定理证明感知器无法做到这一点。这本书非常成功地扼杀了未来 25 年对联结主义的所有资助，这是明斯基感到遗憾的事情，因为他在去世前不久告诉我，他现在意识到了深度神经网络的力量。

In 1969, Minsky wrote his book, Perceptrons, with his colleague, Seymour Papert. The book basically proved a theorem that a perceptron could not devise answers that required the use of the XOR logical function, nor could they solve the connectedness problem. There are two maze-like images on the cover of that book, and if you look carefully, you can see one is fully connected, and the other is not. Making that classification is called the connectedness problem. The theorem proved that a perceptron could not do that. The book was very successful in killing all funding for connectionism for the next 25 years, which is something Minsky regretted, as shortly before he died he told me that he now appreciated the power of deep neural nets.

马丁·福特：马文·明斯基在 50 年代确实研究过早期的联结主义神经网络，对吗？

MARTIN FORD: Marvin Minsky did work on early connectionist neural nets back in the ‘50s, though, right?

雷·库兹韦尔：没错，但到了 20 世纪 60 年代，他对神经网络失去了信心，并没有真正意识到多层神经网络的威力。直到几十年后，当人们尝试使用 3 层神经网络时，这种神经网络的效果才有所改善，人们才意识到这一点。由于梯度爆炸或梯度消失问题，层数过多会产生问题，这基本上是由于数字太大或太小，系数值的动态范围会下降。

RAY KURZWEIL: That’s right, but he became disillusioned with them by the 1960s, and really didn’t appreciate the power of multi-layer neural nets. It was not apparent until decades later when 3-layer neural nets were tried and they worked somewhat better. There was a problem going with too many layers, because of the exploding gradient or vanishing gradient problem, which is basically where the dynamic range of the values of the coefficients would decline because the numbers got too big or too small.

杰弗里·辛顿和一群数学家解决了这个问题，现在我们可以达到任意数量的层级。他们的解决方案是，在每一层之后重新校准信息，这样它就不会超出可以表示的值的范围，这些 100 层的神经网络非常成功。不过，还有一个问题，可以用座右铭来总结，“生命始于十亿个例子。”

Geoffrey Hinton and a group of mathematicians solved that problem and now we can go to any number of levels. Their solution was that you recalibrate the information after each level, so it doesn’t outstrip the range of values that can be represented and these 100-layer neural nets have been very successful. There’s still a problem though, which is summarized by the motto, “Life begins at a billion examples.”

我加入谷歌的原因之一是，我们确实有十亿个示例，比如狗和猫的图片以及其他经过注释的图像类别，但也有很多东西我们没有十亿个示例。我们有很多语言示例，但它们没有注释其含义，而且我们怎么能用我们根本无法理解的语言来注释它们呢？有些问题我们可以解决这个问题，下围棋就是一个很好的例子。DeepMind 系统接受了所有在线动作的训练，数量级为一百万步。这不是十亿。这创造了一个公平的业余玩家，但他们需要另外 9.99 亿个例子，那么他们要从哪里获得这些例子呢？

One of the reasons I’m here at Google is that we do have a billion examples of some things like pictures of dogs and cats and other image categories that are annotated, but there are also lots of things we don’t have a billion examples of. We have lots of examples of language, but they’re not annotated with what they mean, and how could we annotate them anyway using language that we can’t understand in the first place? There’s a certain category of problems where we can work around that, and playing Go is a good example. The DeepMind system was trained on all of the online moves, which is in the order of a million moves. That’s not a billion. That created a fair amateur player, but they need another 999 million examples, so where are they going to get them from?

马丁·福特：你的意思是，目前的深度学习非常依赖标记数据和所谓的监督学习。

MARTIN FORD: What you’re getting at is that deep learning right now is very dependent on labeled data and what’s called supervised learning.

雷·库兹韦尔：是的。解决这个问题的一种方法是，如果你能模拟你工作的世界，那么你就可以创建自己的训练数据，这就是 DeepMind 通过让它自己下棋所做的。他们可以用传统的注释方法来注释这些动作。随后 AlphaZero 实际上训练了一个神经网络来改进注释，因此它能够在没有人类训练数据的情况下以 100 比 0 击败 AlphaGo。

RAY KURZWEIL: Right. One way to work around it is if you can simulate the world you’re working in, then you can create your own training data, and that’s what DeepMind did by having it play itself. They could annotate the moves with traditional annotation methods. Subsequently AlphaZero actually trained a neural net to improve on the annotation, so it was able to defeat AlphaGo 100 games to 0 starting with no human training data.

问题是，在什么情况下可以这样做？例如，我们可以这样做的另一种情况是数学，因为我们可以模拟数学。数论的公理并不比围棋规则更复杂。

The question is, in what situations can you do that in? For example, another situation where we can do that is math, because we can simulate math. The axioms of number theory are no more complicated than the rules of Go.

另一种情况是自动驾驶汽车，尽管驾驶比棋盘游戏或数学系统的公理复杂得多。Waymo 的方法是，通过多种方法组合，创建了一个相当不错的系统，然后驾驶数百万英里，人类驾驶员随时准备接管。这产生了足够的数据来创建一个精确的驾驶世界模拟器。他们现在已经在模拟器中使用模拟车辆行驶了大约十亿英里，这为旨在改进算法的深度神经网络生成了训练数据。尽管驾驶世界比棋盘游戏复杂得多，但这种方法还是奏效了。

Another situation is self-driving cars, even though driving is much more complex than a board game or the axioms of a math system. The way that worked is that Waymo created a pretty good system with a combination of methods and then drove millions of miles with humans at the wheel ready to take over. That generated enough data to create an accurate simulator of the world of driving. They’ve now driven on the order of a billion miles with simulated vehicles in the simulator, which has generated training data for a deep neural net designed to improve the algorithms. This has worked even though the world of driving is much more complex than a board game.

下一个值得尝试模拟的激动人心的领域是生物和医学。如果我们能够模拟生物学（这并非不可能），那么我们就可以在几小时内而不是几年内完成临床试验，并且可以像自动驾驶汽车、棋盘游戏或数学一样生成自己的数据。

The next exciting area to attempt to simulate is the world of biology and medicine. If we could simulate biology, and it’s not impossible, then we could do clinical trials in hours rather than years, and we could generate our own data just like we’re doing with self-driving cars or board games or math.

这不是解决提供足够训练数据问题的唯一方法。人类可以从更少的数据中学习，因为我们参与迁移学习，从可能与我们试图学习的情况截然不同的情况中学习。我有一个不同的学习模型，它基于对人类大脑皮层工作原理的粗略了解。1962 年，我提出了一篇关于人类大脑如何运作的论文，过去 50 年来，我一直在思考思考问题。我的模型不是一个大的神经网络，而是许多小模块，每个模块都可以识别一种模式。在我的书《如何创造思维》中，我将大脑皮层描述为基本上有 3 亿个这样的模块，每个模块都可以识别一个连续的模式并接受一定程度的变化。这些模块按层次结构组织，这是通过它们自己的思考创建的。系统创建了自己的层次结构。

That’s not the only approach to the problem of providing sufficient training data. Humans can learn from much less data because we engage in transfer learning, using learning from situations which may be fairly different from what we are trying to learn. I have a different model of learning based on a rough idea of how the human neocortex works. In 1962 I came up with a thesis on how I thought the human brain works, and I’ve been thinking about thinking for the last 50 years. My model is not one big neural net, but rather many small modules, each of which can recognize a pattern. In my book, How to Create a Mind, I describe the neocortex as basically 300 million of those modules, and each can recognize a sequential pattern and accept a certain amount of variability. The modules are organized in a hierarchy, which is created through their own thinking. The system creates its own hierarchy.

大脑皮层的分层模型可以从更少的数据中学习。人类也是如此。我们可以从少量的数据中学习，因为我们可以将信息从一个领域推广到另一个领域。

That hierarchical model of the neocortex can learn from much less data. It’s the same with humans. We can learn from a small amount of data because we can generalize information from one domain to another.

谷歌联合创始人之一拉里·佩奇 (Larry Page) 很喜欢我《如何创造心智》一文的论文，因此招募我加入谷歌，将那些想法应用于语言理解。

Larry Page, one of the co-founders of Google, liked my thesis in How to Create a Mind and recruited me to Google to apply those ideas to understanding language.

马丁·福特：您有将这些概念应用到谷歌产品的实际例子吗？

MARTIN FORD: Do you have any real-world examples of you applying those concepts to a Google product?

雷·库兹韦尔：Gmail 上的智能回复（为每封电子邮件提供三个回复建议）是我团队使用这种分层系统的应用程序之一。我们刚刚推出了 Talk to Books（ https://books.google.com/talktobooks/），你用自然语言提出问题，系统会在半秒内阅读 10 万本书（也就是 6 亿个句子），然后返回它能从这 6 亿个句子中找到的最佳答案。这一切都基于语义理解，而不是关键字。

RAY KURZWEIL: Smart Reply on Gmail (which provides three suggestions to reply to each email) is one application from my team that uses this hierarchical system. We just introduced Talk to Books (https://books.google.com/talktobooks/), where you ask a question in natural language and the system then reads 100,000 books in a half-second—that’s 600 million sentences—and then returns the best answers that it can find from those 600 million sentences. It’s all based on semantic understanding, not keywords.

在谷歌，我们在自然语言方面取得了进展，语言是大脑皮层的第一个发明。语言是分层的；我们可以通过语言的层次结构相互分享大脑皮层中的分层思想。我认为艾伦·图灵很有先见之明，他以语言为基础制定了图灵测试，因为我认为它确实需要人类思维和人类智慧的全部范围，才能在人类层面上创造和理解语言。

At Google we’re making progress in natural language, and language was the first invention of the neocortex. Language is hierarchical; we can share the hierarchical ideas we have in our neocortex with each other using the hierarchy of language. I think Alan Turing was prescient in basing the Turing test on language because I think it does require the full range of human thinking and human intelligence to create and understand language at human levels.

马丁·福特：你的最终目标是扩展这个想法，真正制造出一台可以通过图灵测试的机器吗？

MARTIN FORD: Is your ultimate objective to extend this idea to actually build a machine that can pass the Turing test?

雷·库兹韦尔：虽然不是所有人都同意我的观点，但我认为，如果图灵测试组织得当，它实际上是一种非常好的人类智能测试。问题是，图灵在 1950 年撰写的简短论文中，只有几段话谈到了图灵测试，而他遗漏了关键要素。例如，他没有描述如何实际进行测试。在实际进行测试时，测试规则非常复杂，但如果计算机要通过有效的图灵测试，我相信它需要具备人类智能的全部范围。以人类的水平理解语言是最终目标。如果人工智能能够做到这一点，它就可以阅读所有文档和书籍，并学习其他一切。我们正在一点一点地实现这一目标。我们可以理解足够的语义，例如，让我们的 Talk to Books 应用程序能够对问题给出合理的答案，但它仍然没有达到人类的水平。我和米奇·卡普尔就此打了个 2 万美元的长线赌注，所得款项将捐给获胜者选择的慈善机构。我说人工智能将在 2029 年之前通过图灵测试，而他说不会。

RAY KURZWEIL: Not everybody agrees with this, but I think the Turing test, if organized correctly, is actually a very good test of human-level intelligence. The issue is that in the brief paper that Turing wrote in 1950, it’s really just a couple of paragraphs that talked about the Turing test, and he left out vital elements. For example, he didn’t describe how to actually go about administering the test. The rules of the test are very complicated when you actually administer it, but if a computer is to pass a valid Turing test, I believe it will need to have the full range of human intelligence. Understanding language at human levels is the ultimate goal. If an AI could do that, it could read all documents and books and learn everything else. We’re getting there a little bit at a time. We can understand enough of the semantics, for example to enable our Talk to Books application to come up with reasonable answers to questions, but it’s still not at human levels. Mitch Kapor and I have a long-range bet on this for $20,000, with the proceeds to go to the charity of the winner’s choice. I’m saying that an AI will pass the Turing test by 2029, whereas he’s saying no.

马丁·福特：你是否同意，图灵测试如果要成为一种有效的智力测试，可能根本不应该有时间限制？仅仅欺骗某人 15 分钟似乎是一种花招。

MARTIN FORD: Would you agree that for the Turing test to be an effective test of intelligence, there probably shouldn’t be a time limit at all? Just tricking someone for 15 minutes seems like a gimmick.

雷·库兹韦尔：当然，如果你看看我和米奇·卡普尔制定的规则，我们给出了几个小时的时间，但也许这还不够。底线是，如果人工智能真的能让你相信它是人类，那么它就通过了测试。我们可以讨论这需要多长时间——如果你有一个经验丰富的法官，可能要几个小时——但我同意，如果时间太短，那么你可能会用一些简单的技巧来蒙混过关。

RAY KURZWEIL: Absolutely, and if you look at the rules that Mitch Kapor and I came up with, we gave a number of hours, and maybe even that’s not enough time. The bottom line is that if an AI is really convincing you that it’s human, then it passes the test. We can debate how long that needs to be—probably several hours if you have a sophisticated judge—but I agree that if the time is too short, then you might get away with simple tricks.

马丁·福特：我认为很容易想象一台智能计算机在假装成人类方面并不擅长，因为它可能是一种外星智能。因此，似乎可以进行一项测试，让每个人都同意该机器是智能的，即使它实际上看起来不像人类。我们可能也希望承认这是一个充分的测试。

MARTIN FORD: I think it’s easy to imagine an intelligent computer that just isn’t very good at pretending to be human because it would be an alien intelligence. So, it seems likely that you could have a test where everyone agreed that the machine was intelligent, even though it didn’t actually seem to be human. And we would probably want to recognize that as an adequate test as well.

雷·库兹韦尔：鲸鱼和章鱼拥有巨大的大脑，它们表现出聪明的行为，但它们显然无法通过图灵测试。说普通话而不说英语的中国人无法通过英语图灵测试，因此有很多方法可以不通过测试而变得聪明。关键的说法是相反的：为了通过测试，你必须聪明。

RAY KURZWEIL: Whales and octopi have large brains and they exhibit intelligent behavior, but they’re obviously not in a position to pass the Turing test. A Chinese person who speaks mandarin and not English would not pass the English Turing test, so there are lots of ways to be intelligent without passing the test. The key statement is the converse: In order to pass the test, you have to be intelligent.

马丁·福特：您是否相信深度学习与您的分层方法相结合真的是前进的方向，或者您是否认为需要进行其他大规模的范式转变才能使我们达到 AGI /人类水平的智能？

MARTIN FORD: Do you believe that deep learning, combined with your hierarchical approach, is really the way forward, or do you think there needs to be some other massive paradigm shift in order to get us to AGI/human-level intelligence?

雷·库兹韦尔：不，我认为人类使用这种分层方法。每个模块都能够进行学习，我实际上在书中指出，大脑中并不是每个模块都进行深度学习，它们所做的是相当于马尔可夫过程的事情，但实际上使用深度学习更好。

RAY KURZWEIL: No, I think humans use this hierarchical approach. Each of these modules is capable of doing learning, and I actually make the case in my book that in the brain they’re not doing deep learning in each module, they’re doing something equivalent to a Markov process, but it actually is better to use deep learning.

在 Google 的系统中，我们使用深度学习来创建表示每个模块中的模式的向量，然后我们有一个超越深度学习范式的层次结构。不过，我认为这对 AGI 来说已经足够了。在我看来，分层方法是人类大脑的工作方式，现在有很多来自大脑逆向工程项目的证据。

In our systems at Google we use deep learning to create vectors that represent the patterns in each module and then we have a hierarchy that goes beyond the deep learning paradigm. I think that’s sufficient for AGI, though. The hierarchical approach is how the human brain does it in my view, and there’s a lot of evidence now for that from the brain reverse engineering projects.

有一种观点认为，人类大​​脑遵循的是规则系统而非联结主义系统。人们指出，人类能够做出清晰的区分，我们能够进行逻辑推理。关键在于联结主义可以模仿基于规则的方法。在特定情况下，联结主义系统可能对其判断非常确定，以至于它看起来和表现得像一个基于规则的系统，但它也能够处理罕见的例外情况及其表面规则的细微差别。

There’s an argument that human brains follow a rule-based system rather than a connectionist one. People point out that humans are capable of having sharp distinctions and we’re capable of doing logic. A key point is that connectionism can emulate a rule-based approach. A connectionist system in a certain situation might be so certain of its judgment that it looks and acts like a rule-based system, but then it’s also able to deal with rare exceptions and the nuances of its apparent rules.

基于规则的系统实际上无法模拟联结系统，因此相反的说法并不成立。Doug Lenat 的“Cyc”是一个令人印象深刻的项目，但我相信它证明了基于规则的系统的局限性。你会达到一个复杂性的上限，规则变得如此复杂，以至于如果你试图修复一件事，你就会破坏另外三件事。

A rule-based system really cannot emulate a connectionist system, so the converse statement is not the case. Doug Lenat’s “Cyc” is an impressive project, but I believe that it proves the limitations of a rule-based system. You reach a complexity ceiling, where the rules get so complex that if you try to fix one thing, you break three other things.

马丁·福特：Cyc 是一个人们尝试手动输入常识逻辑规则的项目吗？

MARTIN FORD: Cyc is the project where people are manually trying to enter logic rules for common sense?

雷·库兹韦尔：是的。我不确定具体数量，但他们制定了海量规则。他们有一种模式，可以打印出行为背后的原因，而解释则会长达数页，很难理解。这是一项令人印象深刻的工作，但它确实表明，这实际上并不是方法，至少不是方法本身，也不是人类实现智能的方式。我们没有遵循一连串的规则，我们有这种分层的自组织方法。

RAY KURZWEIL: Right. I’m not sure of the count, but they have a vast number of rules. They had a mode where it could print out its reasoning for a behavior and the explanations would go on for a number of pages and are very hard to follow. It’s impressive work, but it does show that this is really not the approach, at least not by itself, and it’s not how humans achieve intelligence. We don’t have cascades of rules that we go through, we have this hierarchical self-organizing approach.

我认为分层但联结主义方法的另一个优势是它更善于自我解释，因为你可以查看层次结构中的模块，看看哪个模块影响哪个决策。当你拥有这些庞大的 100 层神经网络时，它们就像一个巨大的黑匣子。很难理解它的推理，尽管已经有人尝试这样做。我确实认为这种分层联结主义方法是一种有效的方法，这就是人类的思维方式。

I think another advantage of a hierarchical, but connectionist approach is that it’s better at explaining itself because you can look at the modules in the hierarchy and see which module influences which decision. When you have these massive 100-layer neural nets, they act like a big black box. It’s very hard to understand its reasoning, though there have been some attempts to do that. I do think that this hierarchical spin on a connectionist approach is an effective approach, and that’s how humans think.

马丁·福特：人类的大脑中有一些结构在出生时就已经存在了。例如，婴儿可以识别面孔。

MARTIN FORD: There are some structures, though, in the human brain, even at birth. For example, babies can recognize faces.

雷·库兹韦尔：我们确实有一些特征生成器。例如，在我们的大脑中，有一个叫做梭状回的模块，它包含专门的电路，可以计算某些比率，比如鼻尖和鼻尾的比率，或者两眼之间的距离。有十几个相当简单的特征，实验表明，如果我们从图像中生成这些特征，然后生成具有相同特征（相同比率）的新图像，那么人们会立即认出它们是同一个人的照片，即使图像中的其他细节已经发生了很大变化。有各种这样的特征生成器，有些带有音频信息，我们计算某些比率并识别部分泛音，然后这些特征输入到分层联结系统中。因此，了解这些特征生成器很重要，在识别人脸方面有一些非常具体的特征，这就是婴儿所依赖的。

RAY KURZWEIL: We do have some feature generators. For example, in our brains we have this module called the fusiform gyrus that contains specialized circuitry and computes certain ratios, like the ratio of the tip of the nose to the end of the nose, or the distance between the eyes. There is set of a dozen or so fairly simple features, and experiments have shown that if we generate those features from images and then generate new images that have the same features—the same ratios—then people will immediately recognize them as a picture of that same person, even though other details have changed quite a bit in the image. There are various feature generators like that, some with audio information that we compute certain ratios and recognize partial overtones, and these features then feed into the hierarchical connectionist system. So, it is important to understand these feature generators, and there are some very specific features in recognizing faces, and that’s what babies rely on.

马丁·福特：我想谈谈通用人工智能（AGI）的发展路径和时机。我认为 AGI 和人类级别的人工智能是等同的。

MARTIN FORD: I’d like to talk about the path and the timing for Artificial General Intelligence (AGI). I’m assuming AGI and human-level AI are equivalent terms.

雷·库兹韦尔：它们是同义词，我不喜欢 AGI 这个词，因为我认为这是对 AI 的含蓄批评。AI 的目标一直是实现越来越高的智能，最终达到人类的智能水平。然而，随着我们的进步，我们已经分离出不同的领域。例如，一旦我们掌握了字符识别，它就变成了 OCR 的独立领域。语音识别和机器人技术也是如此，人们认为 AI 的总体领域不再专注于通用智能。我的观点始终是，我们将通过一次解决一个问题，逐步实现通用智能。

RAY KURZWEIL: They’re synonyms, and I don’t like the term AGI because I think it’s an implicit criticism of AI. The goal of AI has always been to achieve greater and greater intelligence and ultimately to reach human levels of intelligence. As we’ve progressed, though, we’ve spun off separate fields. For example, once we mastered recognizing characters, it became the separate field of OCR. The same happened with speech recognition and robotics, and it was felt that the overarching field of AI was no longer focusing on general intelligence. My view is always that we’ll get to general intelligence step by step by solving one problem at a time.

另一个值得注意的地方是，人类在任何类型的任务中的表现范围都很广。人类在围棋中的表现水平如何？范围很广，从初学围棋的孩子到世界冠军。我们发现，一旦计算机能够达到人类水平，即使是处于该范围的低端，它也会很快超越人类的表现。一年多前，计算机在围棋中的表现还很低，然后它们很快就超越了那个水平。最近，AlphaZero 经过几个小时的训练，就超越了 AlphaGo，以 100 比 0 击败了它。

Another bit of color on that is that human performance in any type of task is a very broad range. What is the human performance level in Go? It’s a broad range from a child who’s playing their first game to the world champion. One thing we’ve seen is that once a computer can achieve human levels, even at the low end of that range, it very quickly soars past human performance. A little over a year ago computers were playing at a low-level in Go and then they quickly soared past that. More recently, AlphaZero soared past AlphaGo and beat it 100 games to 0, after training for a few hours.

计算机的语言理解能力也在提高，但速度并不相同，因为它们还没有足够的现实世界知识。计算机目前无法很好地进行多链推理，基本上是从多个语句中进行推理，同时考虑现实世界的知识。例如，在三年级的语言理解测试中，计算机无法理解，如果一个男孩的鞋子沾满泥巴，他很可能是在外面的泥巴里走路时把鞋子弄脏的，如果他把泥巴弄到厨房的地板上，他的妈妈会很生气。这些对于我们人类来说似乎都是显而易见的，因为我们可能经历过这种情况，但对于人工智能来说却不是那么明显。

Computers are also improving in their language understanding, but not at the same rate, because they don’t yet have sufficient real-world knowledge. Computers currently can’t do multi-chain reasoning very well, basically taking inferences from multiple statements while at the same time considering real-world knowledge. For example, on a third-grade language understanding test, a computer didn’t understand that if a boy had muddy shoes he probably got them muddy by walking in the mud outside and if he got the mud on the kitchen floor it would make his mother mad. That may all seem obvious to us humans because we may have experienced that, but it’s not obvious to the AI.

我认为，从计算机在某些语言测试中表现出来的普通成人理解能力到超越人类的能力，这一过程不会很快实现，因为我认为要做到这一点，还有更多根本性的问题需要解决。尽管如此，正如我们所见，人类的表现范围很广，一旦计算机进入这个范围，它们最终就能超越它，成为超人。它们在语言理解方面的表现达到任何成人水平这一事实非常令人印象深刻，因为我认为语言需要人类智能的全部范围，并且具有人类模糊性和层次化思维的全部范围。总而言之，是的，人工智能正在取得非常迅速的进步，是的，所有这些都在使用联结主义方法。

I don’t think the process will be as quick to go from the average adult comprehension performance that we have now for computers on some language tests to superhuman performance because I think there are more fundamental issues to solve to do that. Nonetheless, human performance is a broad range, as we’ve seen, and once computers get in that range they can ultimately soar past it to become superhuman. The fact that they’re performing at any kind of adult level in language understanding is very impressive because I feel that language requires the full range of human intelligence, and has the full range of human ambiguity and hierarchical thinking. To sum up, yes, AI is making very rapid progress and yes, all of this is using connectionist approaches.

我刚刚和我的团队讨论了除了已经做过的事情之外，我们还需要做些什么才能通过图灵测试。我们已经具备了一定程度的语言理解能力。一个关键要求是多链推理——能够考虑概念的推论和含义——这是重中之重。这是聊天机器人经常失败的一个领域。

I just had a discussion with my team here about what we have to do to pass the Turing test beyond what we’ve already done. We already have some level of language understanding. One key requirement is multi-chain reasoning—being able to consider the inferences and implications of concepts—that’s a high priority. That’s one area where chatbots routinely fail.

如果我说我担心女儿在幼儿园的表现，你不会想在三轮之后问，你有孩子吗？聊天机器人会这样做，因为它们不会考虑所说的一切的所有推论。正如我所提到的，还有现实世界知识的问题，但如果我们能够理解语言的所有含义，那么就可以通过阅读和理解网上的大量文件来获得现实世界的知识。我认为我们对如何做这些事情有很好的想法，而且我们有足够的时间去做。

If I say I’m worried about my daughter’s performance in nursery school, you wouldn’t want to then ask three turns later, do you have any children? Chatbots do that kind of thing because they’re not considering all the inferences of everything that has been said. As I mentioned, there is also the issue of real-world knowledge, but if we could understand all the implications of language, then real-world knowledge could be gained by reading and understanding the many documents available online. I think we have very good ideas on how to do those things and we have plenty of time to do them.

马丁·福特：您长期以来一直直言不讳，您认为人类水平的人工智能将在 2029 年到来。现在仍然如此吗？

MARTIN FORD: You’ve been very straightforward for a long time that the year when you think human-level AI is going to arrive is 2029. Is that still the case?

雷·库兹韦尔：是的。在我 1989 年出版的《智能机器时代》一书中，我设定的预测范围是 2029 年左右，前后 10 年左右。1999 年，我出版了《精神机器时代》，并做出了 2029 年的具体预测。斯坦福大学召开了一次人工智能专家会议，以处理这个显然令人吃惊的预测。当时，我们还没有即时投票机，所以我们基本上是举手表决。当时的共识是，这将需要数百年时间，约四分之一的人说这永远不会发生。

RAY KURZWEIL: Yes. In my book, The Age of Intelligent Machines, which came out in 1989, I put a range around 2029 plus or minus a decade or so. In 1999 I published The Age of Spiritual Machines and made the specific prediction of 2029. Stanford University held a conference of AI experts to deal with this apparently startling prediction. At that time, we didn’t have instant polling machines, so we basically had a show of hands. The consensus view then was it would take hundreds of years, with about a quarter of the group saying it would never happen.

2006 年，达特茅斯学院举办了一次会议，庆祝 1956 年达特茅斯会议 50 周年，我之前提到过，当时我们确实有即时投票设备，共识是 50 年左右。12 年后，也就是 2018 年，现在的共识观点是 20 到 30 年左右，也就是 2038 到 2048 年之间，所以我仍然比人工智能专家的共识更乐观，但只是略微乐观一点。我的观点和人工智能专家的共识越来越接近，但不是因为我改变了我的观点。越来越多的人认为我太保守了。

In 2006 there was a conference at Dartmouth College celebrating the 50th anniversary of the 1956 Dartmouth conference, which I mentioned earlier, and there we did have instant polling devices and the consensus was about 50 years. 12 years later, in 2018 the consensus view now is about 20 to 30 years, so anywhere from 2038 to 2048, so I’m still more optimistic than the consensus of AI experts, but only slightly. My view and the consensus view of AI experts is getting closer together, but not because I’ve changed my view. There’s a growing group of people who think I’m too conservative.

马丁·福特：2029 年距离现在只有 11 年了，其实并不算太远。我有一个 11 岁的女儿，这让我真正意识到了这一点。

MARTIN FORD: 2029 is only 11 years away, which is not that far away really. I have an 11-year-old daughter, which really brings it into focus.

雷·库兹韦尔：进步是指数级的；看看去年的惊人进步。我们在自动驾驶汽车、语言理解、围棋和许多其他领域都取得了巨大进步。无论是硬件还是软件，发展速度都非常快。在硬件方面，指数级增长甚至比一般计算的速度还要快。过去几年，我们每三个月就能将深度学习的可用计算量翻一番，而一般计算量则需要一年才能翻一番。

RAY KURZWEIL: The progress is exponential; look at the startling progress just in the last year. We’ve made dramatic advances in self-driving cars, language understanding, playing Go and many other areas. The pace is very rapid, both in hardware and software. In hardware, the exponential progression is even faster than for computation generally. We have been doubling the available computation for deep learning every three months over the past few years, compared to a doubling time of one year for computation in general.

马丁·福特：一些对人工智能有着深入了解的聪明人仍然预测人工智能的实现需要 100 多年时间。您认为这是因为他们陷入了线性思维的陷阱吗？

MARTIN FORD: Some very smart people with a deep knowledge of AI are still predicting that it will take over 100 years, though. Do you think that is because they are falling into that trap of thinking linearly?

雷·库兹韦尔：A）他们的思维是线性的，B）他们容易受到我所说的工程师悲观主义的影响——他们过于专注于一个问题，觉得这个问题很难解决，因为他们还没有解决这个问题，并推断他们自己能够以他们目前的工作速度解决这个问题。考虑一个领域的进步速度以及思想如何相互作用并将其作为一种现象进行研究是一门完全不同的学科。有些人就是无法理解进步的指数性质，尤其是在信息技术领域。

RAY KURZWEIL: A) they are thinking linearly, and B) they are subject to what I call the engineer’s pessimism—that is being so focused on one problem and feeling that it’s really hard because they haven’t solved it yet, and extrapolating that they alone are going to solve the problem at the pace they’re working on. It’s a whole different discipline to consider the pace of progress in a field and how ideas interact with each other and study that as a phenomenon. Some people are just not able to grasp the exponential nature of progress, particularly when it comes to information technology.

人类基因组计划进行到一半时，7 年后才收集到 1% 的数据，主流批评人士说：“我告诉过你这行不通。7 年内收集 1% 的数据意味着需要 700 年，就像我们说的那样。”我的反应是：“我们完成了 1%——我们快完成了。我们每年都在翻一番。1% 距离 100% 只有 7 倍的翻番。”果然，7 年后它就完成了。

Halfway through the human genome project, 1% had been collected after 7 years, and mainstream critics said, “I told you this wasn’t going to work. 1% in 7 years means it’s going to take 700 years, just like we said.” My reaction was, “We finished one percent—We’re almost done. We’re doubling every year. 1% is only 7 doublings from 100%.” And indeed, it was finished 7 years later.

一个关键问题是，为什么有些人能轻易获得这种能力，而其他人却不能？这绝对不是成就或智力的结果。一些非专业领域的人很容易理解这一点，因为他们只需在智能手机上就能体验到这种进步，而其他非常有成就、处于领域顶尖的人却有这种非常顽固的线性思维。所以，我真的没有答案。

A key question is why do some people readily get this, and other people don’t? It’s definitely not a function of accomplishment or intelligence. Some people who are not in professional fields understand this very readily because they can experience this progress just in their smartphones, and other people who are very accomplished and at the top of their field just have this very stubborn linear thinking. So, I really don’t actually have an answer for that.

马丁·福特：不过，您同意这不仅仅是计算速度或内存容量方面的指数级进步吗？显然，在教会计算机像人类一样从实时、非结构化数据中学习，或者在推理和想象方面，必须实现一些基本的概念突破？

MARTIN FORD: You would agree though that it’s not just about exponential progress in terms of computing speed or memory capacity? There are clearly some fundamental conceptual breakthroughs that have to happen in terms of teaching computers to learn from real time, unstructured data the way that human beings do, or in reasoning and imagination?

雷·库兹韦尔：好吧，软件的发展也是指数级的，尽管它有你提到的不可预测的方面。思想的相互影响本质上是指数级的，一旦我们在一个层面上建立了性能，就会出现进入下一个层面的想法。

RAY KURZWEIL: Well, progress in software is also exponential, even though it has that unpredictable aspect that you’re alluding to. There’s a cross-fertilization of ideas that is inherently exponential, and once we have established performance at one level, ideas emerge to get to the next level.

奥巴马政府科学顾问委员会曾就这个问题做过研究。他们对比了硬件和软件的进步。他们选取了十几个经典的工程技术问题，定量分析了硬件的进步。一般来说，从那时开始的 10 年里，硬件的进步比约为 1,000:1，这与每年性价比翻番的寓意相符。正如你所预料的那样，软件的进步各不相同，但在每种情况下，软件的进步都比硬件大。进步往往是指数级的。如果你在软件方面取得进步，它并不是线性进步，而是指数级进步。总体进步是硬件和软件进步的产物。

There was a study done by the Obama administration scientific advisory board on this question. They examined how hardware and software progress compares. They took a dozen classical engineering and technical problems and looked at the advance quantitatively to see how much was attributable to hardware. Generally, over the previous 10 years from that point, it was about 1,000 to 1 in hardware, which is consistent with the implication of doubling in price performance every year. The software, as you might expect, varied, but in every case, it was greater than the hardware. Advances tend to be exponential. If you make an advance in software, it doesn’t progress linearly; it progresses exponentially. On the overall progress is the product of the progress in hardware and software.

马丁·福特：您预测的另一个日期是 2045 年，也就是您所说的奇点。我认为大多数人会将之与智力爆炸或真正的超级智能的出现联系起来。这样想对吗？

MARTIN FORD: The other date that you’ve given as a projection is 2045 for what you referred to as the singularity. I think most people associate that with an intelligence explosion or the advent of a true superintelligence. Is that the right way to think about it?

雷·库兹韦尔：关于奇点，实际上有两种思想流派：一种是硬起飞学派，一种是软起飞学派。我其实属于软起飞学派，他们认为我们将继续呈指数级增长，这已经够吓人的了。智能爆炸的概念是，在某个神奇的时刻，计算机可以访问自己的设计并对其进行修改，从而创造出更智能的版本，并且它会以非常快速的迭代循环不断这样做，最终智能就会爆发。

RAY KURZWEIL: There are actually two schools of thought on the singularity: there’s a hard take off school and a soft take off school. I’m actually in the soft take off school that says we will continue to progress exponentially, which is daunting enough. The idea of an intelligence explosion is that there is a magic moment where a computer can access its own design and modify it and create a smarter version of itself, and that it keeps doing that in a very fast iterative loop and just explodes in its intelligence.

我认为，自从我们发明技术以来，我们实际上已经这样做了几千年了。技术确实让我们变得更聪明。你的智能手机是大脑的延伸，它确实让我们变得更聪明。这是一个指数级的过程。一千年前，范式转变和进步需要几个世纪的时间，看起来什么都没有发生。你的祖父母和你过着同样的生活，你希望你的孙子也这样做。现在，我们每年都会看到变化，甚至更快。它是指数级的，这导致进步加速，但从这个意义上说，它不是爆炸式的。

I think we’ve actually been doing that for thousands of years, ever since we created technology. We are certainly smarter as a result of our technology. Your smartphone is a brain extender, and it does make us smarter. It’s an exponential process. A thousand years ago paradigm shifts and advances took centuries, and it looked like nothing was happening. Your grandparents lived the same lives you did, and you expected your grandchildren to do the same. Now, we see changes on an annual basis if not faster. It is exponential and that results in an acceleration of progress, but it’s not an explosion in that sense.

我认为到 2029 年，我们的智能将达到人类水平，而且它马上就会成为超人。以我们的 Talk to Books 为例，你问它一个问题，它会在半秒钟内读出 6 亿个句子、10 万本书。就我个人而言，读 10 万本书需要几个小时！

I think we will achieve a human level of intelligence by 2029 and it’s immediately going to be superhuman. Take for example our Talk to Books, you ask it a question and it reads 600 million sentences, 100,000 books, in half a second. Personally, it takes me hours to read 100,000 books!

你的智能手机现在能够根据关键词和其他方法进行搜索，并能非常快速地搜索所有人类知识。谷歌搜索已经超越了关键词搜索，并具备一定的语义能力。语义理解尚未达到人类水平，但比人类思维快十亿倍。软件和硬件都将继续以指数级的速度改进。

Your smartphone right now is able to do searching based on keywords and other methods and search all human knowledge very quickly. Google search already goes beyond keyword search and has some semantic capability. The semantic understanding is not yet at human levels, but it’s a billion times faster than human thinking. And both the software and the hardware will continue to improve at an exponential pace.

马丁·福特：您还因利用技术扩展和延长人类寿命的想法而闻名。您能告诉我更多有关这方面的内容吗？

MARTIN FORD: You’re also well known for your thoughts on using technology to expand and extend human life. Could you let me know more about that?

雷·库兹韦尔：我的一个论点是，我们将与我们正在创造的智能技术融合。我设想的场景是，我们将把医用纳米机器人送入我们的血液中。这些医用纳米机器人的一个应用是扩展我们的免疫系统。这就是我所说的彻底延长寿命的第三座桥梁。第一座桥梁是我们现在可以做的事情，第二座桥梁是完善生物技术和重新编程生命软件。第三座桥梁构成了这些医用纳米机器人，以完善免疫系统。这些机器人还将进入大脑，从神经系统内部而不是通过连接到我们身体外部的设备提供虚拟和增强现实。医用纳米机器人最重要的应用是，我们将把我们的大脑皮层的顶层连接到云端的合成大脑皮层。

RAY KURZWEIL: One thesis of mine is that we’re going to merge with the intelligent technology that we are creating. The scenario that I have is that we will send medical nanorobots into our bloodstream. One application of these medical nanorobots will be to extend our immune systems. That’s what I call the third bridge to radical life extension. The first bridge is what we can do now, and bridge two is the perfecting of biotechnology and reprogramming the software of life. Bridge three constitutes these medical nanorobots to perfect the immune system. These robots will also go into the brain and provide virtual and augmented reality from within the nervous system rather than from devices attached to the outside of our bodies. The most important application of the medical nanorobots is that we will connect the top layers of our neocortex to synthetic neocortex in the cloud.

马丁·福特：这是你在谷歌正在研究的事情吗？

MARTIN FORD: Is this something that you’re working on at Google?

雷·库兹韦尔：我和谷歌团队所做的项目使用了所谓的大脑皮层的粗略模拟。我们尚未完全理解大脑皮层，但我们正在利用现有知识对其进行近似模拟。我们现在能够利用语言实现有趣的应用，但到 2030 年代初，我们将拥有非常好的大脑皮层模拟。

RAY KURZWEIL: The projects I have done with my team here at Google use what I would call crude simulations of the neocortex. We don’t have a perfect understanding of the neocortex yet, but we’re approximating it with the knowledge we have now. We are able to do interesting applications with language now, but by the early 2030s we’ll have very good simulations of the neocortex.

就像你的手机通过访问云端让自己变得聪明一百万倍一样，我们将直接通过大脑做到这一点。这是我们已经通过智能手机做到的事情，尽管它们不在我们的身体和大脑中，我认为这是一种武断的区别。我们用手指、眼睛和耳朵，但它们仍然是大脑的延伸。未来，我们将能够直接通过大脑做到这一点，但不仅仅是直接从大脑执行搜索和语言翻译等任务，而是将我们大脑皮层的顶层与云端的合成大脑皮层连接起来。

Just as your phone makes itself a million times smarter by accessing the cloud, we will do that directly from our brain. It’s something that we already do through our smartphones, even though they’re not inside our bodies and brains, which I think is an arbitrary distinction. We use our fingers and our eyes and ears, but they are nonetheless brain extenders. In the future, we’ll be able to do that directly from our brains, but not just to perform tasks like search and language translation directly from our brains, but to actually connect the top layers of our neocortex to synthetic neocortex in the cloud.

两百万年前，我们的额头还没有这么大，但随着我们的进化，我们的额头变得更大，可以容纳更多的新皮层。我们用它做了什么？我们把它放在了新皮层层次的顶端。作为灵长类动物，我们已经做得很好了，现在我们能够以更抽象的水平思考。

Two million years ago, we didn’t have these large foreheads, but as we evolved we got a bigger enclosure to accommodate more neocortex. What did we do with that? We put it at the top of the neocortical hierarchy. We were already doing a very good job at being primates, and now we were able to think at an even more abstract level.

这是我们发明技术、科学、语言和音乐的推动因素。我们发现的每种人类文化都有音乐，但没有灵长类文化有音乐。现在这是一次性交易，我们不能继续扩大围栏，因为生育将变得不可能。两百万年前的这种新皮层扩张实际上使生育变得相当困难。

That was the enabling factor for us to invent technology, science, language, and music. Every human culture that we have discovered has music, but no primate culture has music. Now that was a one-shot deal, we couldn’t keep growing the enclosure because birth would have become impossible. This neocortical expansion two million years ago actually made birth pretty difficult as it was.

2030 年代大脑皮层的这种新扩展不会是一次性的事情。就在我们说话的时候，云的威力每年都在翻倍。它不受固定外壳的限制，所以我们思维的非生物部分将继续增长。如果我们算一下，到 2045 年，我们的智力将增加 10 亿倍，这是一个如此深刻的转变，以至于很难看到事件视界之外的东西。所以，我们借用了物理学中事件视界的比喻，以及看到它之外的东西的难度。

This new extension in the 2030s to our neocortex will not be a one-shot deal. Even as we speak, the cloud is doubling in power every year. It’s not limited by a fixed enclosure, so the non-biological portion of our thinking will continue to grow. If we do the math, we will multiply our intelligence a billion-fold by 2045, and that’s such a profound transformation that it’s hard to see beyond that event horizon. So, we’ve borrowed this metaphor from physics of the event horizon and the difficulty of seeing beyond it.

谷歌搜索和 Talk to Books 等技术的速度至少比人类快 10 亿倍。它还没有达到人类的智能水平，但一旦我们达到这一点，人工智能将利用已经存在的巨大速度优势以及容量和能力的持续指数增长。这就是奇点的意义，它是一个软起飞，但指数增长仍然变得相当令人生畏。如果你把某样东西翻倍 30 倍，你就乘以了 10 亿。

Technologies such as Google Search and Talk to Books are at least a billion times faster than humans. It’s not at human levels of intelligence yet, but once we get to that point, AI will take advantage of the enormous speed advantage which already exists and an ongoing exponential increase in capacity and capability. So that’s the meaning of the singularity, it’s a soft take off, but exponentials nonetheless become quite daunting. If you double something 30 times, you’re multiplying by a billion.

马丁·福特：您经常谈到奇点的影响领域之一是医学，尤其是人类寿命，这也许是您受到批评的一个领域。我听过您去年在麻省理工学院的一次演讲，您在演讲中说，在 10 年内，大多数人可能能够达到您所说的“长寿逃逸速度”，您还说您个人可能已经实现了这一目标？您真的相信它会这么快发生吗？

MARTIN FORD: One of the areas where you’ve talked a lot about the singularity having an impact is in medicine and especially in the longevity of human life, and this is maybe one area where you’ve been criticized. I heard a presentation you gave at MIT last year where you said that within 10 years, most people might be able to achieve what you call “longevity escape velocity,” and you also said that you think you personally might have achieved that already? Do you really believe it could happen that soon?

雷·库兹韦尔：就生物技术而言，我们正处于一个转折点。人们看待医学时，认为医学只会像过去一样，以同样不靠谱的速度缓慢发展。医学研究基本上是靠谱的。制药公司会从一份包含数千种化合物的清单中寻找有一定影响力的药物，而不是真正理解和系统地重新编程生命软件。

RAY KURZWEIL: We are now at a tipping point in terms of biotechnology. People look at medicine, and they assume that it is just going to plod along at the same hit or miss pace that they have been used to in the past. Medical research has essentially been hit or miss. Drug companies will go through a list of several thousand compounds to find something that has some impact, as opposed to actually understanding and systematically reprogramming the software of life.

说我们的基因过程是软件，这不仅仅是一个比喻。它是一串数据，它进化于一个时代，因为食物等资源有限，所以每个个体活得太久不符合人类的利益。我们正在从一个稀缺时代转变为一个富足时代

It’s not just a metaphor to say that our genetic processes are software. It is a string of data, and it evolved in an era where it was not in the interest of the human species for each individual to live very long because there were limited resources such as food. We are transforming from an era of scarcity to an era of abundance

生物学作为信息处理过程，其各个方面的能力每年都在翻番。例如，基因测序就是这样。第一个基因组花费了 10 亿美元，现在我们花费了近 1,000 美元。但我们不仅能够收集这种原始的生命目标代码，而且能够理解它、对其进行建模、模拟它，最重要的是对其进行重新编程，这种能力每年也在翻番。

Every aspect of biology as an information process has doubled in power every year. For example, genetic sequencing has done that. The first genome cost US $1 billion, and now we’re close to $1,000. But our ability to not only collect this raw object code of life but to understand it, to model it, to simulate it, and most importantly to reprogram it, is also doubling in power every year.

我们现在正在获得临床应用——虽然现在只是涓涓细流，但未来十年将会如洪水般涌来。数百项重大干预措施正在通过监管渠道。我们现在可以修复心脏病发作后破碎的心脏，也就是说，使用重新编程的成体干细胞，使心脏病发作后射血分数较低的心脏恢复活力。我们可以培育器官，并成功地将它们植入灵长类动物体内。免疫疗法基本上就是重新编程免疫系统。免疫系统本身并不能对抗癌症，因为它并没有进化出对抗我们日后患上的疾病的能力。我们实际上可以重新编程它，让它识别癌症并将其作为病原体进行治疗。这是癌症治疗的一大亮点，而且有令人惊奇的试验，几乎每个参加试验的人都从 4 期晚期癌症转为缓解期。

We’re now getting clinical applications—it’s a trickle today, but it’ll be a flood over the next decade. There are hundreds of profound interventions in process that are working their way through the regulatory pipeline. We can now fix a broken heart from a heart attack, that is, rejuvenate a heart with a low ejection fraction after a heart attack using reprogrammed adult stem cells. We can grow organs and are installing them successfully in primates. Immunotherapy is basically reprogramming the immune system. On its own, the immune system does not go against cancer because it did not evolve to go after diseases that tend to get us later on in life. We can actually reprogram it and turn it on to recognize cancer and treat it as a pathogen. This is a huge bright spot in cancer treatment, and there are remarkable trials where virtually every person in the trial goes from stage 4 terminal cancer to being in remission.

十年后，医学将发生翻天覆地的变化。如果你勤奋努力，我相信你将能够达到长寿逃逸速度，这意味着我们将比过去延长更多的时间，不仅是婴儿的预期寿命，还有你的剩余预期寿命。这并不是保证，因为你明天仍然可能被所谓的公共汽车撞到，而预期寿命实际上是一个复杂的统计概念，但时间之沙将开始流逝而不是流逝。再过十年，我们也将能够逆转衰老过程。

Medicine is going to be profoundly different in a decade from now. If you’re diligent, I believe you will be able to achieve longevity escape velocity, which means that we’ll be adding more time than is going by, not just to infant life expectancy but to your remaining life expectancy. It’s not a guarantee, because you can still be hit by the proverbial bus tomorrow, and life expectancy is actually a complicated statistical concept, but the sands of time will start running in rather than running out. In another decade further out, we’ll be able to reverse aging processes as well.

马丁·福特：我想谈谈人工智能的缺点和风险。我想说，有时你会受到不公平的批评，被批评为对这一切过于乐观，甚至有点过于乐观。在这些发展方面，我们有什么需要担心的吗？

MARTIN FORD: I want to talk about the downsides and the risks of AI. I would say that sometimes you are unfairly criticized as being overly optimistic, maybe even a bit Pollyannaish, about all of this. Is there anything we should worry about in terms of these developments?

雷·库兹韦尔：我写过关于 GNR 弊端的文章比任何人都多，而这比史蒂芬·霍金或伊隆·马斯克表达他们的担忧早了几十年。我于1999 年出版的《精神机器时代》一书中对 GNR（遗传学、纳米技术和机器人技术，即人工智能）的弊端进行了广泛的讨论，这促使比尔·乔伊在 2000 年 1 月撰写了他​​著名的《连线》封面故事，题为《为什么未来不需要我们》。

RAY KURZWEIL: I’ve written more about the downsides than anyone, and this was decades before Stephen Hawking or Elon Musk were expressing their concerns. There was extensive discussion of the downsides of GNR—Genetics, Nanotechnology, and Robotics (which means AI)—in my book, The Age of Spiritual Machines, which came out in 1999 that led Bill Joy to write his famous Wired cover story in January 2000 titled, Why the Future Doesn’t Need Us.

马丁·福特：这句话是根据大学炸弹客泰德·卡辛斯基的一句话说出来的，不是吗？

MARTIN FORD: That was based upon a quote from Ted Kaczynski, the Unabomber, wasn’t it?

雷·库兹韦尔：我在其中一页引用了他的话，听起来像是在表达一种非常冷静的担忧，然后你翻过一页，你会看到这句话来自《大学炸弹客宣言》。我在那本书中相当详细地讨论了 GNR 的生存风险。在我 2005 年出版的《奇点临近》一书中，我详细探讨了 GNR 风险这一主题。第 8 章的标题是“GNR 的深层交织的承诺与危险。”

RAY KURZWEIL: I have a quote from him on one page that sounds like a very level-headed expression of concern, and then you turn the page, and you see that this is from the Unabomber Manifesto. I discussed in quite some detail in that book the existential risk of GNR. In my 2005 book, The Singularity is Near, I go into the topic of GNR risks in a lot of detail. Chapter 8 is titled, “The Deeply Intertwined Promise versus Peril of GNR.”

我乐观地认为，人类将度过难关。我们从技术中获得的益处远大于危害，但你不必看得太远就能看到技术带来的深远危害，例如，20 世纪的所有破坏——尽管 20 世纪实际上是当时最和平的世纪，而我们现在正处于一个更加和平的时代。世界正在变得越来越好，例如，过去 200 年里贫困减少了 95%，全球识字率从不到 10% 上升到 90% 以上。

I’m optimistic that we’ll make it through as a species. We get far more profound benefit than harm from technology, but you don’t have to look very far to see the profound harm that has manifested itself, for example, in all of the destruction in the 20th century—even though the 20th century was actually the most peaceful century up to that time, and we’re in a far more peaceful time now. The world is getting profoundly better, for example, poverty has been cut 95% in the last 200 years and literacy rates have gone from under 10% to over 90% in the world.

人们判断世界是变好还是变坏的算法是“我多久听到一次好消息和坏消息？”，而这不是一个很好的方法。曾有一项对 26 个国家的 24,000 人进行的民意调查询问了这个问题：“过去 20 年，全球贫困状况是变好了还是变坏了？”87% 的人错误地认为贫困状况在恶化。只有 1% 的人正确地认为过去 20 年里贫困率下降了一半或更多。人类在进化过程中偏爱坏消息。一万年前，关注坏消息非常重要，例如，树叶沙沙作响的声音可能意味着有捕食者在靠近。关注坏消息比研究你的庄稼比去年高出了半个百分点更重要，而我们仍然偏爱坏消息。

People’s algorithm for whether the world is getting better or worse is “how often do I hear good news versus bad news?”, and that’s not a very good method. There was a poll taken of 24,000 people in about 26 countries asking this question, “Is poverty worldwide getting better or worse over the last 20 years?” 87% said, incorrectly, that it’s getting worse. Only 1% said correctly that it’s fallen by half or more in the last 20 years. Humans have an evolutionary preference for bad news. 10,000 years ago, it was very important that you paid attention to bad news, for example that little rustling in the leaves that might be a predator. That was more important to pay attention to than studying that your crops are half a percent better than last year, and we continue to have this preference for bad news.

马丁·福特：然而，实际风险和生存风险之间存在着一个阶跃变化。

MARTIN FORD: There’s a step-change, though, between real risks and existential risks.

雷·库兹韦尔：我们在应对信息技术带来的生存风险方面也做得相当不错。四十年前，一群有远见的科学家看到了生物技术的前景和危险，当时生物技术还未成熟，他们召开了第一届阿西洛马生物技术伦理大会。这些伦理标准和策略定期更新。效果很好。因故意或意外滥用或生物技术问题而受到伤害的人数已接近于零。我们现在开始获得我提到的巨大利益，未来十年，这种利益将如洪流般涌现。

RAY KURZWEIL: Well, we’ve also done reasonably well with existential risks from information technology. Forty years ago, a group of visionary scientists saw both the promise and the peril of biotechnology, neither of which was close at hand at the time, and they held the first Asilomar Conference on biotechnology ethics. These ethical standards and strategies have been updated on a regular basis. That has worked very well. The number of people who have been harmed by intentional or accidental abuse or problems with biotechnology has been close to zero. We’re now beginning to get the profound benefit that I alluded to, and that’s going to become a flood over the next decade.

这是这种综合道德标准和确保技术安全的技术策略的成功，其中大部分现已纳入法律。这并不意味着我们可以将生物技术带来的危险从我们的担忧清单中划掉；我们不断推出更强大的技术，如 CRISPR，我们必须不断重新制定标准。

That’s a success for this approach of comprehensive ethical standards, and technical strategies on how to keep the technology safe, and much of that is now baked into law. That doesn’t mean we can cross danger from biotechnology off our list of concerns; we keep coming up with more powerful technologies like CRISPR and we have to keep reinventing the standards.

大约 18 个月前，我们召开了第一届阿西洛马人工智能伦理会议，会上我们制定了一套道德标准。我认为这些标准需要进一步完善，但总体而言，这是一个可行的方法。我们必须高度重视它。

We had our first AI ethics Asilomar conference about 18 months ago where we came up with a set of ethical standards. I think they need further development, but it’s an overall approach that can work. We have to give it a high priority.

马丁·福特：目前真正引起人们关注的问题是所谓的控制问题或协调问题，超级智能的目标可能与人类的最佳利益不一致。你会认真对待这个问题吗？是否应该努力解决这个问题？

MARTIN FORD: The concern that’s really getting a lot of attention right now is what’s called the control problem or the alignment problem, where a superintelligence might not have goals that are aligned with what’s best for humanity. Do you take that seriously, and should work be done on that?

雷·库兹韦尔：人类的目标并不完全一致，而这才是关键问题。将人工智能视为一种独立的文明是一种误解，就好像它是来自火星的外星人入侵一样。我们创造工具来扩大我们自己的触角。10,000 年前，我们无法在更高的树枝上获取食物，所以我们制造了一种工具来扩大我们的触角。我们无法用赤手建造摩天大楼，所以我们有了利用我们肌肉力量的机器。非洲一个拥有智能手机的孩子只需敲击几下键盘，就能获得所有人类知识。

RAY KURZWEIL: Humans don’t all have aligned goals with each other, and that’s really the key issue. It’s a misconception to talk about AI as a civilization apart, as if it’s an alien invasion from Mars. We create tools to extend our own reach. We couldn’t reach food at that higher branch 10,000 years ago, so we made a tool that extended our reach. We can’t build a skyscraper with our bare hands, so we have machines that leverage the range of our muscles. A kid in Africa with a smartphone is connected to all of the human knowledge with a few keystrokes.

这就是技术的作用；它使我们能够超越自己的局限，而这正是我们正在做的，并且将继续利用人工智能去做的。这不是我们与人工智能的对抗，这一直是许多人工智能未来主义反乌托邦电影的主题。我们将与人工智能融合。我们已经这样做了。你的手机并不在你的身体和大脑中，这一事实是一个没有区别的区别，因为它可能就在你的身体和大脑中。没有手机，我们就不会出门，没有手机，我们就不完整，今天，没有手机，没有人能够工作、接受教育或维持人际关系，而且我们与手机的联系越来越密切。

That is the role of technology; it enables us to go beyond our limitations, and that’s what we are doing and will continue to do with AI. It’s not us versus the AIs, which has been the theme of many AI futurist dystopian movies. We are going to merge with it. We already have. The fact that your phone is not physically inside your body and brain is a distinction without a difference, because it may as well be. We don’t leave home without it, we’re incomplete without it, nobody could do their work, get their education, or keep their relationships without their devices today, and we’re getting more intimate with them.

我之所以选择麻省理工学院，是因为它在 1965 年就已经非常先进，拥有一台计算机。当时我必须骑着自行车穿过校园才能到达那里，并出示身份证才能进入大楼。而现在，半个世纪过去了，我们把计算机放在口袋里，并且一直在使用它们。它们已经融入我们的生活，最终将融入我们的身体和大脑。

I went to MIT because it was so advanced in 1965 that it had a computer. I had to take my bicycle across the campus to get to it and show my ID to get into the building, and now half a century later we’re carrying them in our pockets, and we’re using them constantly. They are integrated into our lives and will ultimately become integrated into our bodies and brains.

如果你看看几千年来我们经历的冲突和战争，你会发现它们都是由于人类之间存在分歧而引起的。我确实认为技术实际上往往会创造更大的和谐、和平和民主化。你可以将民主化的兴起追溯到通信的改进。两个世纪前，世界上只有一个民主国家。一个世纪前有六个民主国家。现在，在 192 个公认的国家中，有 123 个是民主国家，占世界总数的 64%。世界并不是一个完美的民主国家，但民主实际上已被接受为当今的标准。这是人类历史上最和平的时代，生活的方方面面都在变得更好，这是由于技术的影响，它变得越来越智能，并且深深融入了我们的生活。

If you look at the conflict and warfare we’ve had over the millennia, it’s been from humans having disagreements. I do think technology tends to actually create greater harmony and peace and democratization. You can trace the rise of democratization to improvements in communication. Two centuries ago, there was only one democracy in the world. There were half a dozen democracies one century ago. Now there are 123 democracies out of 192 recognized countries, that’s 64% of the world. The world’s not a perfect democracy, but democracy has actually been accepted as the standard today. It is the most peaceful time in human history, and every aspect of life is getting better, and this is due to the effect of technology which is becoming increasingly intelligent, and it’s deeply integrated into who we are.

如今，不同人类群体之间存在着冲突，而每个群体之间的冲突都因技术而加剧。这种情况将继续存在，尽管我认为还有另一个主题，即更好的通信技术可以充分利用我们的短距离同理心。我们对小群体有生理上的同理心，但现在我们能够真正体验地球另一边的人所发生的事情，这种同理心被放大了。我认为这是关键问题；在我们通过技术增强个人能力的同时，我们仍然必须管理好人际关系。

We have conflict today between different groups of humans, each of whom are amplified by their technology. That will continue to be the case, although I think there’s this other theme that better communication technology harnesses our short-range empathy. We have a biological empathy for small groups of people, but that’s now amplified by our ability to actually experience what happens to people half a world away. I think that’s the key issue; we still have to manage our human relations as we increase our personal powers through technology.

马丁·福特：我们来谈谈经济和就业市场混乱的可能性。我个人认为，失业或技能下降以及不平等现象加剧的可能性很大。实际上，我认为这可能会造成新工业革命规模的混乱。

MARTIN FORD: Let’s talk about the potential for economic and job market disruption. I personally do think there’s a lot of potential for jobs to be lost or deskilled and for greatly increasing inequality. I actually think it could be something that will be disruptive on the scale of a new Industrial Revolution.

雷·库兹韦尔：我想问您一个问题：上一次工业革命的结果如何？两百年前，织布工人享受着代代相传的行业协会，这种行业协会已经传承了数百年。当这些纺线和织布机问世，他们的生计被彻底颠覆时，他们的商业模式被彻底颠覆。他们预测，更多的机器将会出现，大多数人将失去工作，只有精英才能享受就业机会。这一预测部分成真——更多的纺织机被引入，许多技能和工作被淘汰。然而，随着社会变得更加繁荣，就业率不但没有下降，反而上升了。

RAY KURZWEIL: Let me ask you this: how did that last Industrial Revolution work out? Two hundred years ago, the weavers had enjoyed a guild that was passed down from generation to generation for hundreds of years. Their business model was turned on its head and disrupted when all these thread-spinning and cloth-weaving machines came out that completely upended their livelihoods. They predicted that more machines would come out and that most people would lose their jobs, and that employment would be enjoyed just by an elite. Part of that prediction came true—more textile machines were introduced and many types of skills and jobs were eliminated. However, employment went up, not down as society became more prosperous.

如果我是 1900 年一位有先见之明的未来学家，我会指出你们当中 38% 的人在农场工作，25% 的人在工厂工作，但我预测 115 年后，即 2015 年，农场工作人数将达到 2%，工厂工作人数将达到 9%。每个人的反应都会是，“天哪，我要失业了！”然后我会说“别担心，被淘汰的工作都是技能阶梯最底层的工作，我们将在技能阶梯的顶端创造更多工作。”

If I were a prescient futurist in 1900 I would point out that 38% of you work on farms and 25% of you work in factories, but I predict that 115 years from now, in 2015, that’ll be 2% on farms, and 9% in factories. Everybody’s reaction would be, “Oh my god I’m going to be out of work!” I would then say “Don’t worry, the jobs that are eliminated are at the bottom of the skill ladder, and we are going to create an even larger number of jobs at the top of the skill ladder.”

人们会说，“哦，真的吗？什么新工作？”我会说，“嗯，我不知道，我们还没有发明它们。”人们说我们摧毁的工作比我们创造的多得多，但事实并非如此，我们的工作岗位从 1900 年的 2400 万个增加到今天的 1.42 亿个，占人口的比例从 31% 增加到 44%。这些新工作相比如何？首先，今天的平均工作时薪以不变美元计算比 1900 年高 11 倍。因此，我们把工作时间从大约 3,000 小时缩短到 1,800 小时。人们每年的收入仍然是不变美元的 6 倍，而且工作也变得更加有趣。我认为，即使在下一次工业革命中，这种情况仍将继续。

People would say, “Oh really, what new jobs?”, and I’d say, “Well I don’t know, we haven’t invented them yet.” People say we’ve destroyed many more jobs than we’ve created but that’s not true, we’ve gone from 24 million jobs in 1900 to 142 million jobs today, and as a percentage of the population that goes from 31% to 44%. How do these new jobs compare? Well, for one thing, the average job today pays 11 times as much in constant dollars per hour than in 1900. As a result, we’ve shortened the work year from about 3,000 hours to 1,800 hours. People still make 6 times as much per year in constant dollars, and the jobs have become much more interesting. I think that’s going to continue to be the case even in the next Industrial Revolution.

马丁·福特：真正的问题是这次是否不同。你对之前发生的事情的看法当然是正确的，但根据大多数估计，劳动力中可能有一半或更多的人正在做基本上可预测和相对常规的工作，所有这些工作都可能受到机器学习的威胁。自动化大多数可预测的工作并不需要人类级别的人工智能。

MARTIN FORD: The real question is whether this time it’s different. What you say about what happened previously is certainly true, but it is also true, according to most estimates, that maybe half or more of the people in the workforce are doing things that are fundamentally predictable and relatively routine, and all those jobs are going to be potentially threatened by machine learning. Automating most of those predictable jobs does not require human-level AI.

机器人工程师、深度学习研究人员等可能会创造出新的工作岗位，但你不可能指望现在所有翻汉堡或开出租车的人都从事这些工作，即使假设这些新工作岗位数量足够多。我们谈论的是一种可以取代人们认知、取代脑力的技术，而且它的影响范围将极其广泛。

There may be new kinds of work created for robotics engineers and deep learning researchers and all of that, but you cannot take all the people that are now flipping hamburgers or driving taxis and realistically expect to transition them into those kinds of jobs, even assuming that there are going to be a sufficient number of these new jobs. We’re talking about a technology that can displace people cognitively, displace their brainpower, and it’s going to be extraordinarily broad-based.

雷·库兹韦尔：你的预测中隐含的模型是我们与他们的对抗，人类将如何与机器对抗。我们已经让自己变得更聪明，以便完成这些更高级的工作。我们让自己变得更聪明不是通过直接连接到我们大脑的东西，而是通过智能设备。没有这些大脑扩展器，任何人都无法完成工作，大脑扩展器将进一步扩展我们的大脑，它们将更紧密地融入我们的生活。

RAY KURZWEIL: Your model that’s implicit in your prediction is us-versus-them, and what are the humans going to do versus the machines. We’ve already made ourselves smarter in order to do these higher-level types of jobs. We’ve made ourselves smarter not with things connected directly into our brains yet, but with intelligent devices. Nobody can do their jobs without these brain extenders, and the brain extenders are going to extend our brains even further, and they’re going to be more closely integrated into our lives.

我们为提高技能所做的一件事就是教育。1870 年，我们有 68,000 名大学生，而今天，我们有 1500 万名大学生。如果把他们和为他们服务的所有人（如教职员工）算上，大约有 20% 的劳动力只参与高等教育，而且我们不断创造新事物。大约六年前，整个应用经济还不存在，而它构成了当今经济的主要组成部分。我们要让自己变得更聪明。

One thing that we did to improve our skills is education. We had 68,000 college students in 1870 and today we have 15 million. If you take them and all the people that service them, such as faculty and staff, it is about 20 percent of the workforce that is just involved in higher education, and we are constantly creating new things to do. The whole app economy did not exist about six years ago, and that forms a major part of the economy today. We’re going to make ourselves smarter.

在考虑这个问题时，需要考虑的另一个论点是我之前提到的激进富足论。在国际货币基金组织年度会议上，我与国际货币基金组织总裁克里斯蒂娜·拉加德进行了一次台上对话，她说：“与此相关的经济增长在哪里？数字世界拥有这些奇妙的东西，但从根本上讲，你不能吃信息技术，不能穿它，也不能生活在其中，”我的回答是，“所有这些都将改变。”

A whole other thesis that needs to be looked at in considering this question is the radical abundance thesis that I mentioned earlier. I had an on-stage dialogue with Christine Lagarde, the managing director of the IMF, at the annual International Monetary Fund meeting and she said, “Where’s the economic growth associated with this? The digital world has these fantastic things, but fundamentally you can’t eat information technology, you can’t wear it, you can’t live in it,” and my response was, “All that’s going to change.”

“所有这些名义上的实体产品都将成为信息技术。我们将在人工智能控制的建筑中通过垂直农业种植粮食，种植水培水果和蔬菜，并通过体外克隆肌肉组织来获取肉类，以极低的成本提供不含化学物质的高品质食品，并且不会让动物遭受痛苦。信息技术的通货紧缩率为 50%；你只需花一半的价格就能获得一年前可以购买的相同计算、通信和基因测序，而这种大规模的通货紧缩将影响这些传统的实体产品。”

“All those types of nominally physical products are going to become an information technology. We’re going to grow food with vertical agriculture in AI-controlled buildings with hydroponic fruits and vegetables, and in vitro cloning of muscle tissue for meat, providing very high-quality food without chemicals at very low cost, and without animal suffering. Information technology has a 50% deflation rate; you get the same computation, communication, genetic sequencing that you could purchase a year ago for half the price, and this massive deflation is going to attend to these traditionally physical products.”

马丁·福特：那么，您认为 3D 打印、机器人工厂和农业等技术可以降低几乎所有产品的成本吗？

MARTIN FORD: So, you think that technologies like 3D printing or robotic factories and agriculture could drive costs down for nearly everything?

雷·库兹韦尔：没错，3D 打印将在 2020 年代打印出衣服。由于各种原因，我们还没有完全实现这一目标，但一切都在朝着正确的方向发展。我们需要的其他实物也将通过 3D 打印机打印出来，包括几天内就能组装成一栋建筑的模块。我们所需要的所有实物最终都将由这些人工智能控制的信息技术来实现。

RAY KURZWEIL: Exactly, 3D printing will print out clothing in the 2020s. We’re not quite there yet for various reasons, but all that’s moving in the right direction. The other physical things that we need will be printed out on 3D printers, including modules which will snap together a building in a matter of days. All the physical things we need will ultimately become facilitated by these AI-controlled information technologies.

通过应用深度学习来开发更好的材料，太阳能得到了进一步的发展，因此，能源存储和收集的成本正在迅速下降。太阳能总量每两年翻一番，风能也呈现出同样的趋势。可再生能源现在只需翻五倍，即每两年翻一番，就能满足我们 100% 的能源需求，届时它将使用来自太阳或风能的千分之一的能量。

Solar energy is being facilitated by applying deep learning to come up with better materials, and as a result, the cost of both energy storage and energy collection is coming down rapidly. The total amount of solar energy is doubling every two years, and the same trend exists with wind energy. Renewable energy is now only about five doublings, at two years per doubling, away from meeting 100% of our energy needs, by which time it will use one part in thousands of the energy from the sun or from the wind.

克里斯蒂娜·拉加德说：“好吧，有一种资源永远不会成为信息技术，那就是土地。我们已经挤在一起了。”我回答说：“那只是因为我们决定挤在一起，创建城市，这样我们就可以一起工作和娱乐。”随着我们的虚拟通信变得更加强大，人们已经开始分散。试着坐火车去世界上任何地方，你会发现 95% 的土地都未被利用。

Christine Lagarde said, “OK, there is one resource that will never be an information technology, and that’s land. We are already crowded together.” I responded “That’s only because we decided to crowd ourselves together and create cities so we could work and play together.” People are already spreading out as our virtual communication becomes more robust. Try taking a train trip anywhere in the world and you will see that 95% of the land is unused.

到 2030 年代，我们将能够为所有人、为整个人类提供极高的生活质量，这种生活质量将超越我们今天所认为的高生活水平。我在 TED 上预测，到 2030 年代，我们将实现全民基本收入，实际上不需要那么多钱就能提供极高的生活水平。

We’re going be able to provide a very high quality of living that’s beyond what we consider a high standard of living today for everyone, for all of the human population, as we get to the 2030s. I made a prediction at TED that we will have universal basic income, which won’t actually need to be that much to provide a very high standard of living, as we get into the 2030s.

马丁·福特：所以，你最终会成为基本收入的支持者吗？你同意不会有适合所有人的工作，或者也许不是每个人都需要工作，人们会有一些其他的收入来源，比如全民基本收入？

MARTIN FORD: So, you’re a proponent of a basic income, eventually? You agree that there won’t be a job for everyone, or maybe everyone won’t need a job, and that there’ll be some other source of income for people, like a universal basic income?

雷·库兹韦尔：我们认为工作是通往幸福的途径。我认为关键问题在于目的和意义。人们仍然会竞争，以便能够做出贡献并获得满足感。

RAY KURZWEIL: We assume that a job is a road to happiness. I think the key issue will be purpose and meaning. People will still compete to be able to contribute and get gratification.

马丁·福特：但你不一定要为你从中获得意义的事情获得报酬，对吗？

MARTIN FORD: But you don’t necessarily have to get paid for the thing that you get meaning from?

雷·库兹韦尔：我认为我们将改变经济模式，而且我们已经在这样做了。我的意思是，上大学被认为是一件值得做的事情。这不是一份工作，但它被认为是一项有价值的活动。你不需要工作收入就能拥有满足物质生活需求的良好生活水平，我们将继续沿着马斯洛需求层次上升。我们一直在这样做，只要将今天与 1900 年进行比较即可。

RAY KURZWEIL: I think we will change the economic model and we are already in the process of doing that. I mean, being a student in college is considered a worthwhile thing to do. It’s not a job, but it’s considered a worthwhile activity. You won’t need income from a job in order to have a very good standard of living for the physical requirements of life, and we will continue to move up Maslow’s hierarchy. We have been doing that, just compare today to 1900.

马丁·福特：您如何看待与中国在先进人工智能领域的竞争？中国在隐私等方面的监管较少，确实具有优势。此外，他们的人口数量更多，可以产生更多数据，也意味着他们可能拥有更多年轻的图灵或冯·诺依曼。

MARTIN FORD: What do you think about the perceived competition with China to get to advanced AI? China does have advantages in terms of having less regulation on things like privacy. Plus, their population is so much larger, which generates more data and also means they potentially have a lot more young Turings or von Neumanns in the pipeline.

雷·库兹韦尔：我不认为这是一场零和游戏。如果中国工程师在太阳能或深度学习方面取得突破，我们所有人都会从中受益。中国和美国一样，出版了大量书籍，而且信息实际上被广泛分享。看看谷歌，它将其 TensorFlow 深度学习框架放到了公共领域，我们小组也这样做了，将 Talk to Books 和 Smart Reply 的底层技术开源，以便人们可以使用它。

RAY KURZWEIL: I don’t think it’s a zero-sum game. An engineer in China who comes up with a breakthrough in solar energy or in deep learning benefits all of us. China is publishing a lot just as the United States is, and the information is actually shared pretty widely. Look at Google, which put its TensorFlow deep learning framework into the public domain, and we did that in our group with the technology underlying Talk to Books and Smart Reply being made open source so people can use that.

我个人对中国强调经济发展和创业精神表示欢迎。我最近在中国时，创业精神的蓬勃发展显而易见。我鼓励中国朝着信息自由交流的方向发展。我认为这是这种进步的基础。全世界都把硅谷视为激励模式。硅谷实际上只是创业精神的隐喻，是庆祝实验、将失败视为经验的地方。我认为这是一件好事，我真的不认为这是一场国际竞争。

I personally welcome the fact that China is emphasizing economic development and entrepreneurship. When I was in China recently the tremendous explosion of entrepreneurship was apparent. I would encourage China to move in the direction of free exchange of information. I think that’s fundamental for this type of progress. All around the world we see Silicon Valley as a motivating model. Silicon Valley really is just a metaphor for entrepreneurship, the celebrating of experimenting, and calling failure experience. I think that’s a good thing, I really don’t see it as an international competition.

马丁·福特：但是，你是否担心中国是一个独裁国家，这些技术确实有军事用途？谷歌和伦敦的 DeepMind 等公司已经明确表示，他们不希望自己的技术用于任何与军事有关的领域。而中国的腾讯和百度等公司实际上没有选择权。我们是否应该担心未来会出现某种不对称现象？

MARTIN FORD: But do you worry about the fact that China is an authoritarian state, and that these technologies do have, for example, military applications? Companies like Google and certainly DeepMind in London have been very clear that they don’t want their technology used in anything that is even remotely military. Companies like Tencent and Baidu in China don’t really have the option to make that choice. Is that something we should worry about, that there’s a kind of asymmetry going forward?

雷·库兹韦尔：军事用途与专制政府结构不同。我对中国政府的专制倾向感到担忧，我鼓励他们向更大的信息自由和民主治理方式迈进。我认为这将对他们和每个人的经济都有帮助。

RAY KURZWEIL: Military use is a different issue from authoritarian government structure. I am concerned about the authoritarian orientation of the Chinese government, and I would encourage them to move toward greater freedom of information and democratic ways of governing. I think that will help them and everyone economically.

我认为这些政治、社会和哲学问题仍然非常重要。我担心的不是人工智能会独立发展，因为我认为它与我们紧密结合。我担心的是人类的未来，人类已经是一个技术文明。我们将继续通过技术来提升自己，因此确保人工智能安全的最佳方法是关注我们如何管理自己。

I think these political and social and philosophical issues remain very important. My concern is not that AI is going to go off and do something on its own, because I think it’s deeply integrated with us. I’m concerned about the future of the human population, which is already a human technological civilization. We’re going to continue to enhance ourselves through technology, and so the best way to assure the safety of AI is to attend to how we govern ourselves as humans.

雷·库兹韦尔 被公认为世界最杰出的发明家和未来学家之一。雷在麻省理工学院获得工程学位，师从人工智能领域的奠基人之一马文·明斯基。他后来在多个领域做出了重大贡献。他是第一台 CCD 平板扫描仪、第一台全字体光学字符识别器、第一台盲人用文字转语音阅读机、第一台文本转语音合成器、第一台能够重现大钢琴和其他管弦乐器的音乐合成器以及第一台商业化销售的大词汇量语音识别器的主要发明者。

RAY KURZWEIL is widely recognized as one of the world’s foremost inventors and futurists. Ray received his engineering degree from MIT, where he was mentored by Marvin Minsky, one of the founding fathers of the field of artificial intelligence. He went on to make major contributions in a variety of areas. He was the principal inventor of the first CCD flat-bed scanner, the first omni-font optical character recognition, the first print-to-speech reading machine for the blind, the first text-to-speech synthesizer, the first music synthesizer capable of recreating the grand piano and other orchestral instruments, and the first commercially marketed large-vocabulary speech recognition.

雷获得的荣誉很多，其中包括因在音乐技术领域的杰出成就而荣获的格莱美奖；他是美国国家技术奖章（美国科技领域最高荣誉）获得者，被选入美国国家发明家名人堂，拥有二十一个荣誉博士学位，并获得过三任美国总统的嘉奖。

Among Ray’s many honors, he received a Grammy Award for outstanding achievements in music technology; he is the recipient of the National Medal of Technology (the nation’s highest honor in technology), was inducted into the National Inventors Hall of Fame, holds twenty-one honorary doctorates, and honors from three US presidents.

雷撰写了五本全国畅销书，包括《纽约时报》畅销书《奇点临近》（2005 年）和《如何创造思维》（2012 年）。他是奇点大学的联合创始人兼校长，也是谷歌的工程总监，领导着一个开发机器智能和自然语言理解的团队。

Ray has written five national best-selling books, including New York Times bestsellers The Singularity Is Near (2005) and How To Create A Mind (2012). He is Co-Founder and Chancellor of Singularity University and a Director of Engineering at Google, heading up a team developing machine intelligence and natural language understanding.

雷因其在技术指数级进步方面的研究而闻名，他将其正式称为“加速回报定律”。几十年来，他做出了许多重要的预测，这些预测都被证明是准确的。

Ray is known for his work on exponential progress in technology, which he has formalized as “The Law of Accelerating Returns.” Over the course of decades, he has made a number of important predictions that have proven to be accurate.

雷的第一部小说《丹妮尔，女超人编年史》将于 2019 年初出版。雷的另一本书《奇点越来越近》预计将于 2019 年底出版。

Ray’s first novel, Danielle, Chronicles of a Superheroine, is being published in early 2019. Another book by Ray, The Singularity is Nearer, is expected to be published in late 2019.

麻省理工学院计算机科学与人工智能实验室主任

DIRECTOR OF MIT CSAIL

Daniela Rus 是麻省理工学院计算机科学与人工智能实验室 (CSAIL) 主任，该实验室是世界上最大的人工智能和机器人研究机构之一。Daniela 是 ACM、AAAI 和 IEEE 的院士，也是美国国家工程院和美国艺术与科学学院的成员。Daniela 领导机器人、移动计算和数据科学领域的研究。

Daniela Rus is the Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT, one of the world’s largest research organizations focused on AI and robotics. Daniela is a fellow of ACM, AAAI and IEEE, and a member of the National Academy of Engineering, and the American Academy for Arts and Science. Daniela leads research in robotics, mobile computing, and data science.

马丁·福特：首先让我们谈谈你的背景，看看你是如何对人工智能和机器人产生兴趣的。

MARTIN FORD: Let’s start by talking about your background and looking at how you became interested in AI and robotics.

DANIELA RUS：我一直对科学和科幻小说很感兴趣，小时候我读过当时所有流行的科幻小说。我在罗马尼亚长大，那里的媒体种类并不像美国那么多，但有一部电视剧我非常喜欢，那就是原版《迷失太空》。

DANIELA RUS: I’ve always been interested in science and science fiction, and when I was a kid I read all the popular science fiction books at the time. I grew up in Romania where we didn’t have the range of media that you had in the US, but there was one show that I really enjoyed, and that’s the original Lost in Space.

马丁·福特：我记得。你不是我交谈过的第一个从科幻小说中汲取职业灵感的人。

MARTIN FORD: I remember that. You’re not the first person I’ve spoken to who has drawn their career inspiration from science fiction.

DANIELA RUS：我从未错过《迷失太空》的任何一集，而且我喜欢那个酷酷的怪孩子威尔和机器人。我当时没有想到我会做任何与此有关的事情。我很幸运，数学和科学都很好，到了上大学的年龄，我知道自己想做一些与数学有关的事情，但不是纯数学，因为它似乎太抽象了。我主修计算机科学和数学，辅修天文学——天文学继续与我对其他世界的幻想联系在一起。

DANIELA RUS: I never missed an episode of Lost in Space, and I loved the cool geeky kid Will and the robot. I didn’t imagine that I would do anything remotely associated with that at that time. I was lucky enough to be quite good at math and science, and by the time I got to college age I knew that I wanted to do something with math, but not pure math because it seemed too abstract. I studied computer science with a major in computer science and mathematics, and a minor in astronomy—the astronomy continuing the connection to my fantasies of what could be in other worlds.

在我本科学位即将结束的时候，我听了图灵奖得主、理论计算机科学家约翰·霍普克罗夫特的一次演讲，在那次演讲中，约翰说，经典计算机科学已经完蛋了。他的意思是，计算机领域的创始人提出的许多图论算法都有解决方案，现在是伟大应用的时候了，在他看来，这些应用就是机器人。

Toward the end of my undergraduate degree I went to a talk given by John Hopcroft, the Turing Award-winning theoretical computer scientist, and in that talk, John said that classical computer science was finished. What he meant by that was that many of the graph-theoretic algorithms that were posed by the founders of the field of computing had solutions and it was time for the grand applications, which in his opinion were robots.

我觉得这是一个令人兴奋的想法，所以我和 John Hopcroft 一起攻读博士学位，因为我想为机器人领域做出贡献。然而，当时机器人领域根本没有发展。例如，我们唯一可用的机器人是一个巨大的 PUMA 手臂（可编程通用操作臂），这是一种工业操作器，与我童年时对机器人的幻想几乎没有共同之处。这让我思考了很多我可以做出的贡献，最后我研究了灵巧的操作，但主要是从理论、计算的角度。我记得在完成论文后，我试图实现我的算法，超越模拟并创建真实的系统。不幸的是，当时可用的系统是犹他州/麻省理工学院的手和索尔兹伯里手，这两种手都无法施加我的算法所需的力和扭矩。

I found that an exciting idea, so I worked on my PhD with John Hopcroft because I wanted to make contributions to the field of robotics. However, at that time the field of robotics was not at all developed. For example, the only robot that was available to us was a big PUMA arm (Programmable Universal Manipulation Arm), an industrial manipulator that had little in common with my childhood fantasies of what robots should be. It got me thinking a lot about what I could contribute, and I ended up studying dexterous manipulation, but very much from a theoretical, computational point of view. I remember finishing my thesis and trying to implement my algorithms to go beyond simulation and create real systems. Unfortunately, the systems that were available at the time were the Utah/MIT hand and the Salisbury hand, and neither one of those hands was able to exert the kind of forces and torques that my algorithms required.

马丁·福特：在我看来，物理机器和算法之间存在很大差距。

MARTIN FORD: It sounds to me like there was a big gap between where the physical machines were and where the algorithms were.

DANIELA RUS：没错。当时我才真正意识到，机器实际上是身体和大脑之间的一种紧密联系，对于任何你想让机器执行的任务，你都需要一个能够完成这些任务的身体，然后你需要一个大脑来控制身体，让它完成它应该做的事情。

DANIELA RUS: Exactly. At the time, I really realized that a machine is actually a closed connection between body and brain, and for any task you want that machine to execute, you really needed a body capable of those tasks, and then you needed a brain to control the body to deliver what it was meant to do.

因此，我对身体和大脑之间的互动产生了浓厚的兴趣，并开始质疑机器人的概念。工业机械手是机器人的绝佳例子，但它们并不是我们能用机器人做的所有事情；还有许多其他方式来设想机器人。

As a result, I became very interested in the interaction between body and brain, and challenging the notion of what a robot is. So industrial manipulators are excellent examples of robots, but they are not all that we could do with robots; there are so many other ways to envision robots.

如今，我的实验室里有各种非常非传统的机器人。有模块化细胞机器人、软机器人、食物机器人，甚至还有纸质机器人。我们正在研究新型材料、新型形状、新型架构以及想象机器人身体的不同方式。我们还对这些身体如何运作的数学基础进行了大量研究，我对理解和推进自主科学和智能工程非常感兴趣。

Today in my lab, we have all kinds of very non-traditional robots. There are modular cellular robots, soft robots, robots built out of food, and even robots built out of paper. We’re looking at new types of materials, new types of shapes, new types of architectures and different ways of imagining what the machine body ought to be. We also do a lot of work on the mathematical foundations of how those bodies operate, and I’m very interested in understanding and advancing the engineering of both the science of autonomy and of intelligence.

我对设备硬件和控制硬件的算法之间的联系非常感兴趣。当我思考算法时，我认为虽然考虑解决方案非常重要，但考虑这些解决方案的数学基础也很重要，因为从某种意义上说，这是我们创造知识的宝库，其他人可以在此基础上继续发展。

I became very interested in the connection between the hardware of the device and the algorithms that control the hardware. When I think about algorithms, I think that while it’s very important to consider the solutions, it’s also important to consider the mathematical foundations for those solutions because that’s in some sense where we create the nuggets of knowledge that other people can build on.

马丁·福特：您是麻省理工学院计算机科学与人工智能实验室 (CSAIL) 的主任，该实验室不仅是机器人领域最重要的研究机构之一，也是整个人工智能领域最重要的研究机构之一。您能解释一下 CSAIL 到底是什么吗？

MARTIN FORD: You’re the director of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), which is one of the most important research endeavors in not just robotics, but in AI generally. Could you explain what exactly CSAIL is?

DANIELA RUS：CSAIL 的目标是发明计算的未来，通过计算让世界变得更美好，并培养出一些世界上最优秀的研究学生。

DANIELA RUS: Our objective at CSAIL is to invent the future of computing to make the world better through computing, and to educate some of the best students in the world in research.

CSAIL 是一个非凡的组织。当我还是学生时，我把它视为科技界的奥林匹斯山，从未想过我会成为其中的一员。我喜欢把 CSAIL 视为计算未来的先知，也是人们设想如何利用计算让世界变得更美好的地方。

CSAIL is an extraordinary organization. When I was a student, I looked up to it as the Mount Olympus of technology and never imagined that I’d become a part of it. I like to think of CSAIL as the prophet for the future of computing, and the place where people envision how computing can be used to make the world better.

CSAIL 实际上有两个部分，计算机科学 (CS) 和人工智能，两者都有着非常深厚的历史。我们组织的人工智能方面可以追溯到 1956 年，当时该领域被发明和创立。1956 年，马文·明斯基 (Marvin Minsky) 召集他的朋友们在新罕布什尔州度过了一个月，毫无疑问，他们在树林里徒步旅行，喝酒，进行了精彩的交谈，不受社交媒体、电子邮件和智能手机的干扰。

CSAIL actually has two parts, Computer Science (CS) and AI, both having a really deep history. The AI side of our organization goes back to 1956 when the field was invented and founded. In 1956, Marvin Minsky gathered his friends in New Hampshire where they spent a month, no doubt hiking in the woods, drinking wine and having great conversations, uninterrupted by social media, email, and smartphones.

当他们走出森林时，他们告诉世界，他们创造了一个新的研究领域：人工智能。人工智能是指创造机器的科学和工程，这些机器在感知世界、在世界中移动、玩游戏、推理、交流甚至学习方面都表现出人类水平的技能。我们 CSAIL 的研究人员一直在思考这些问题，并做出了开创性的贡献，能够成为这个社区的一员是一种非凡的荣幸。

When they emerged from the woods, they told the world that they had coined a new field of study: artificial intelligence. AI refers to the science and engineering of creating machines that exhibit human-level skills in how they perceive the world; in how they move in the world; in how they play games; in how they reason; in how they communicate; and even, in how they learn. Our researchers at CSAIL have been thinking about these questions and making groundbreaking contributions ever since, and it’s an extraordinary privilege to be part of this community.

计算机科学方面可以追溯到 1963 年，当时计算机科学家、麻省理工学院教授 Bob Fano 提出了一个疯狂的想法，即两个人可以同时使用同一台计算机。你必须明白，这在当时是一个很大的梦想，当时计算机只有房间那么大，你必须预订使用时间。最初，它被设立为 MAC 项目，代表机器辅助认知，但有一个笑话说，它实际上是以明斯基和 Corby（Fernando “Corby” Corbató）的名字命名的 MAC，他们是计算机科学和人工智能方面的两位技术负责人。自 1963 年实验室成立以来，我们的研究人员投入了大量精力来想象计算是什么样子以及它能完成什么。

The computer science side goes back to 1963, when Bob Fano, a computer scientist and MIT professor, had the crazy idea that two people might use the same computer at the same time. You have to understand this was a big dream back then when computers were the size of rooms and you had to book time on them. Originally, it was set up as Project MAC, which stood for Machine-Aided Cognition, but there was a joke that it was actually named MAC after Minsky and Corby (Fernando “Corby” Corbató), who were the two technical leads for the CS and the AI side. Ever since the founding of the laboratory in 1963, our researchers have put a lot of effort into imagining what computing looks like and what it can accomplish.

如今，很多你认为理所当然的事情都源于 CSAIL 的研究成果，例如密码、RSA 加密、启发 Unix 的计算机分时系统、光学鼠标、面向对象编程、语音系统、具有计算机视觉的移动机器人、自由软件运动等等。最近，CSAIL 在定义云和云计算、通过大规模开放在线课程 (MOOC) 实现教育民主化以及思考安全性、隐私性和计算的许多其他方面方面都处于领先地位。

Many of the things that you take for granted today have their roots in the research developed at CSAIL, such as the password, RSA encryption, the computer time-sharing systems that inspired Unix, the optical mouse, object-oriented programming, speech systems, mobile robots with computer vision, the free software movement, the list goes on. More recently CSAIL has been a leader in defining the cloud and cloud computing, and in democratizing education through Massive Open Online Courses (MOOCs) and in thinking about security, privacy, and many other aspects of computing.

马丁·福特：CSAIL 现在规模有多大？

MARTIN FORD: How big is CSAIL today?

DANIELA RUS：CSAIL 是麻省理工学院最大的研究实验室，拥有 1,000 多名成员，横跨 5 个学院和 11 个部门。CSAIL 目前拥有 115 名教职员工，每位教职员工都对计算抱有远大的梦想，计算是我们精神的重要组成部分。我们的一些教职员工希望通过算法、系统或网络让计算变得更好，而另一些教职员工则希望通过计算让人类的生活更美好。例如，Shafi Goldwasser 希望确保我们可以通过互联网进行私人对话；而 Tim Berners-Lee 希望创建一份权利法案，即万维网的大宪章。我们的研究人员希望确保如果我们生病了，我们可以获得的治疗是个性化和定制的，以尽可能有效。我们的研究人员希望推动机器的功能：Leslie Kaelbling 希望制造出 Data 中尉指挥官，而 Russ Tedrake 则希望制造能够飞行的机器人。我想制造变形机器人，因为我希望看到一个无处不在的机器人世界，它们可以为我们的认知和体力任务提供支持。

DANIELA RUS: CSAIL is the largest research laboratory at MIT, with over 1,000 members, and it cuts across 5 schools and 11 departments. CSAIL today has 115 faculty members, and each of these faculty members has a big dream about computing, which is such an important part of our ethos here. Some of our faculty members want to make computing better through algorithms, systems or networks, while others want to make life better for humanity with computing. For example, Shafi Goldwasser wants to make sure that we can have private conversations over the internet; and Tim Berners-Lee wants to create a bill of rights, a Magna Carta of the World Wide Web. We have researchers who want to make sure that if we get sick, the treatments that are available to us are personalized and customized to be as effective as they can be. We have researchers who want to advance what machines can do: Leslie Kaelbling wants to make Lieutenant-Commander Data, and Russ Tedrake wants to make robots that can fly. I want to make shape-shifting robots because I want to see a world with pervasive robots that support us in our cognitive and physical tasks.

这一愿望的真正灵感来自回顾历史，我们观察到，仅仅 20 年前，计算还是一项只有少数专家才能完成的任务，因为计算机体积庞大、价格昂贵且难以操作，而且需要一定的知识才能知道如何使用它们。十年前，随着智能手机、云计算和社交媒体的出现，这一切都发生了变化。

This aspiration is really inspired by looking back at history and observing that only 20 years ago, computation was a task reserved for the expert few because computers were large, expensive, and difficult to handle, and it took knowledge to know what to do with them. All of that changed a decade ago when smartphones, cloud computing, and social media came along.

如今，很多人都在进行计算。你不必成为专家就可以使用计算，而且你使用计算的频率如此之高，以至于你甚至不知道自己对它的依赖程度有多高。试想一下，如果没有万维网和所有相关功能，你的生活会是怎样的一天。没有社交媒体；没有电子邮件交流；没有 GPS；没有医院诊断；没有数字媒体；没有数字音乐；没有网上购物。计算已经渗透到生活的各个方面，这真是令人难以置信。对我来说，这提出了一个非常令人兴奋且重要的问题：在这个被计算改变的世界里，如果机器人和认知助手帮助我们完成体力和认知任务，情况会是什么样子？

Today, so many people compute. You don’t have to be an expert in order to use computing, and you use computing so much that you don’t even know how much you depend on it. Try to imagine a day in your life without the world wide web and everything that enables. No social media; no communication through email; no GPS; no diagnosis in hospitals; no digital media; no digital music; no online shopping. It’s just incredible to see how computation has permeated into the fabric of life. To me, this raises a very exciting and important question, which is: In this world that has been so changed by computation, what might it look like with robots and cognitive assistants helping us with physical and cognitive tasks?

马丁·福特：作为一家大学组织，您认为纯研究和商业化研究（最终发展成产品）之间的平衡点在哪里？您是成立初创公司还是与商业公司合作？

MARTIN FORD: As a university-based organization, what’s the balance between what you would classify as pure research and things that are more commercial and that end up actually developing into products? Do you spin off startups or work with commercial companies?

DANIELA RUS：我们不为企业提供资助，而是专注于培训学生，为他们毕业后提供各种选择，无论是继续学术生涯、进入高科技行业还是成为企业家。我们全力支持所有这些道路。例如，假设一名学生经过几年的研究创建了一种新型系统，突然间该系统立即有了应用。这就是我们所倡导的技术创业，数百家企业从 CSAIL 研究中脱颖而出，但这些企业并没有得到 CSAIL 的资助。

DANIELA RUS: We don’t house companies; instead, we focus on training our students and giving them various options for what they could do when they graduate, whether that be joining the academic life, going into high-tech industry, or becoming entrepreneurs. We fully support all of those paths. For example, say a student creates a new type of system after several years of research, and all of a sudden there is an immediate application for the system. This is the kind of technological entrepreneurship that we embrace, and hundreds of companies have been spun out of CSAIL research, but the actual companies do not get housed by CSAIL.

我们也不创造产品，但这并不意味着我们忽视它们。我们对我们的工作如何转化为产品感到非常兴奋，但总的来说，我们的使命实际上是着眼于未来。我们考虑的是 5 到 10 年后的问题，这也是我们大部分工作所处的地方，但我们也接受当今重要的想法。

We also don’t create products, but that’s not to say we ignore them. We’re very excited about how our work could be turned into products, but generally, our mission is really to focus on the future. We think about problems that are 5 to 10 years out, and that’s where most of our work is, but we also embrace the ideas that matter today.

马丁·福特：我们来谈谈机器人技术的未来吧，这听起来像是你花了很多时间思考的事情。未来的创新会是怎样的呢？

MARTIN FORD: Let’s talk about the future of robotics, which sounds like something you spend a great deal of your time thinking about. What’s coming down the line in terms of future innovations?

DANIELA RUS：机器人技术已经改变了我们的世界。如今，医生可以与患者沟通，老师可以与千里之外的学生沟通。我们有机器人帮助工厂车间打包，我们有联网传感器用于监控设施，我们有 3D 打印技术可以制造定制商品。人工智能和机器人技术的进步已经改变了我们的世界，当我们考虑从我们的人工智能和机器人系统中添加更广泛的功能时，非凡的事情将成为可能。

DANIELA RUS: Our world has already been transformed by robotics. Today, doctors can connect with patients, and teachers can connect with students that are thousands of miles away. We have robots that help with packing on factory floors, we’ve got networked sensors that we deploy to monitor facilities, and we have 3D printing that creates customized goods. Our world has already been transformed by advances in artificial intelligence and robotics, and when we consider adding even more extensive capabilities from our AI and robot systems, extraordinary things will be possible.

从高层次来看，我们必须想象一个世界，在那里，我们不再需要做常规任务，因为这是当今技术发展的最佳状态。这些常规任务可能是体力任务，也可能是计算或认知任务。

At the high level, we have to picture a world where routine tasks will be taken off our plate because this is the sweet spot for where technology is today. These routine tasks could be physical tasks or could be computational or cognitive tasks.

你已经看到了机器学习在各个行业应用的兴起，但我想想象一个世界，在那里，更多平凡的日常任务将从你的工作中解脱出来。也许垃圾桶可以自动倒出，智能基础设施可以确保垃圾消失，或者机器人可以帮你折叠衣物。我们将拥有像水或电一样的交通工具，你可以随时去任何地方。我们将拥有智能助手，让我们能够最大限度地利用工作时间，优化我们的生活，让我们生活得更好、更健康，工作更有效率。这将是非凡的。

You already see some of that in the rise of machine learning applications for various industries, but I like to think of a world where more mundane routine tasks are taken off your plate. Maybe garbage cans that take themselves out and smart infrastructure to ensure that they disappear, or robots that will fold your laundry. We will have transportation available in the same way that water or electricity are available, and you will be able to go anywhere at any time. We will have intelligent assistants who will enable us to maximize our time at work and optimize our lives to live better and more healthily, and to work more efficiently. It will be extraordinary.

马丁·福特：自动驾驶汽车怎么样？我什么时候才能在曼哈顿叫一辆机器人出租车，让它载我去任何地方？

MARTIN FORD: What about self-driving cars? When will I be able to call a robot taxi in Manhattan and have it take me anywhere?

DANIELA RUS：我要补充一点，现在某些自动驾驶技术已经可用。今天的解决方案适用于某些 4 级自动驾驶情况（汽车工程师协会定义的完全自动驾驶之前的倒数第二级）。我们已经有了可以运送人员和包裹的机器人汽车，它们可以在低复杂度、低交互的环境中以低速行驶。曼哈顿是一个具有挑战性的案例，因为曼哈顿的交通非常混乱，但我们已经有了可以在退休社区或商业园区或交通不太繁忙的一般地方运行的机器人汽车。尽管如此，这些仍然是现实世界中可以遇到其他交通、其他人和其他车辆的地方。

DANIELA RUS: I’m going to qualify my answer and say that certain autonomous driving technologies are available right now. Today’s solutions are good for certain level 4 autonomy situations (the penultimate level before full autonomy, as defined by the Society of Automotive Engineers). We already have robot cars that can deliver people and packages, and that operate at low speeds in low-complexity environments where you have low interaction. Manhattan is a challenging case because traffic in Manhattan is super chaotic, but we do already have robot cars that could operate in retirement communities or business campuses, or in general places where there is not too much traffic. Nevertheless, those are still real-world places where you can expect other traffic, other people, and other vehicles.

接下来，我们必须考虑如何扩展这种能力，使其适用于更大、更复杂的环境，在这些环境中，您将在更高的速度下面临更复杂的交互。这项技术正在慢慢到来，但仍面临一些严峻的挑战。例如，我们今天在自动驾驶中使用的传感器在恶劣天气下不太可靠。我们还有很长的路要走才能达到 5 级自动驾驶，即汽车在任何天气条件下都能完全自动驾驶。这些系统还必须能够处理纽约市的交通拥堵，我们必须更好地将机器人汽车与人类驾驶的汽车结合起来。这就是为什么考虑混合的人机环境非常令人兴奋和非常重要。每年我们都会看到技术的逐步改进，但如果我估计，要得到一个完整的解决方案可能还需要十年。

Next, we have to think about how we extend this capability to make it applicable to bigger and more complex environments where you’ll face more complex interactions at higher speeds. That technology is slowly coming, but there are still some serious challenges ahead. For instance, the sensors that we use in autonomous driving today are not very reliable in bad weather. We’ve still got a long way to go to reach level 5 autonomy where the car is fully autonomous in all weather conditions. These systems also have to be able to handle the kind of congestion that you find in New York City, and we have to become much better at integrating robot cars with human-driven cars. This is why thinking about mixed human/machine environments is very exciting and very important. Every year we see gradual improvements in the technology but getting to a complete solution, if I was to estimate, could take another decade.

不过，在某些特定应用中，我们将比其他应用更早地看到自动驾驶技术的商业化应用。我相信退休社区现在就可以使用自动驾驶班车。我相信自动驾驶卡车的长途驾驶很快就会实现。这比在纽约开车要简单一些，但比在退休社区开车要难一些，因为你必须高速行驶，而且有很多极端情况和情况可能需要人类驾驶员介入。假设下着倾盆大雨，你正处在落基山脉危险的山口。为了应对这种情况，你需要一个非常棒的传感器和控制系统与人类的推理和控制能力进行协作。随着高速公路上的自动驾驶，我们将看到自动驾驶与人类辅助交织在一起，反之亦然，这肯定不会超过 10 年，也许 5 年。

There are, though, specific applications where we will see autonomy used commercially sooner than other applications. I believe that a retirement community could use autonomous shuttles today. I believe that long-distance driving with autonomous trucks is coming soon. It’s a little bit simpler than driving in New York, but it’s harder than what driving in a retirement community would look like because you have to drive at high speed, and there are a lot of corner cases and situations where maybe a human driver would have to step in. Let’s say it’s raining torrentially and you are on a treacherous mountain pass in the Rockies. To face that, you need to have a collaboration between a really great sensor and control system, and the human’s reasoning and control capabilities. With autonomous driving on highways, we will see patches of autonomy interleaved with human assistance, or vice versa, and that will be sooner than 10 years, for sure, maybe 5.

马丁·福特：那么在未来十年内，这些问题中的很多都会得到解决，但不是全部。也许这项服务将仅限于特定的路线或地图绘制得非常清晰的区域？

MARTIN FORD: So in the next decade a lot of these problems would be solved, but not all of them. Maybe the service would be confined to specified routes or areas that are really well mapped?

DANIELA RUS：嗯，不一定。进展正在发生。我们小组刚刚发表了一篇论文，展示了首批能够在乡村道路上行驶的系统之一。因此，一方面，挑战令人生畏，但另一方面，10 年是一段很长的时间。20 年前，施乐 PARC 首席科学家 Mark Weiser 谈到了普适计算，他被视为梦想家。今天，我们已经为他设想的所有使用计算的情况以及计算将如何支持我们的情况提供了解决方案。

DANIELA RUS: Well, not necessarily. Progress is happening. In our group we just released a paper that demonstrates one of the first systems capable of driving on country roads. So, on the one hand, the challenges are daunting, but on the other hand, 10 years is a long time. 20 years ago, Mark Weiser, who was the chief scientist at Xerox PARC, talked about pervasive computing and he was seen as a dreamer. Today, we have solutions for all of the situations he envisioned where computing would be used, and how computing would support us.

我想成为一名技术乐观主义者。我想说，我认为技术具有巨大的潜力，可以团结人们而不是分裂人们，赋予人们力量而不是疏远人们。然而，为了实现这一目标，我们必须推进科学和工程，使技术更强大、更易于部署。

I want to be a technology optimist. I want to say that I see technology as something that has the huge potential to unite people rather than divide people, and to empower people rather than estrange people. In order to get there, though, we have to advance science and engineering to make technology more capable and more deployable.

我们还必须实施能够广泛普及教育的计划，让人们熟悉技术，以便能够充分利用技术，让每个人都能梦想着如何利用技术改善生活。但如今的人工智能和机器人技术还无法做到这一点，因为解决方案需要大多数人不具备的专业知识。我们需要重新审视教育人们的方式，确保每个人都拥有利用技术的工具和技能。我们能做的另一件事是继续发展技术，让机器开始适应人类，而不是人类适应机器。

We also have to embrace programs that enable broad education and allow people to become familiar with technology to the point where they can take advantage of it and where anyone could dream about how their lives could be better by the use of technology. That’s something that’s not possible with AI and robotics today because the solutions require expertise that most people don’t have. We need to revisit how we educate people to ensure that everyone has the tools and the skills to take advantage of technology. The other thing that we can do is to continue to develop the technology side so that machines begin to adapt to people, rather than the other way around.

马丁·福特：对于能够真正做有用的事情的无处不在的个人机器人而言，我认为限制因素实际上是灵活性。老生常谈的是能够让机器人去冰箱给你拿一瓶啤酒。就我们今天的技术而言，这是一个真正的挑战。

MARTIN FORD: In terms of ubiquitous personal robots that can actually do useful things, it seems to me that the limiting factor is really dexterity. The cliché is being able to ask a robot to go to the refrigerator and get you a beer. That’s a real challenge in terms of the technology that we have today.

DANIELA RUS：是的，我认为你说得对。我们目前确实看到导航方面的成功比操控方面要大得多，而这两者是机器人的两大主要能力。导航方面的进步得益于硬件的进步。当 LIDAR 传感器（激光扫描仪）推出后，突然之间，无法与声纳配合使用的算法开始发挥作用，这是一次变革。我们现在有了一种可靠的传感器，控制算法可以以稳健的方式使用它。因此，地图绘制、规划和定位开始兴起，这激发了人们对自动驾驶的极大热情。

DANIELA RUS: Yes, I think you’re right. We do currently see significantly greater successes in navigation than in manipulation, and these are two major types of capabilities for robots. The advances in navigation were enabled by hardware advances. When the LIDAR sensor—the laser scanner—was introduced, all of a sudden, the algorithms that didn’t work with sonar started working, and that was transformational. We now had a reliable sensor that control algorithms could use in a robust way. As a result of that, mapping, planning, and localization took off, and that fueled the great enthusiasm in autonomous driving.

回到灵活性的问题，在硬件方面，我们的大多数机械手看起来仍然像 50 年前一样。我们的大多数机械手仍然非常僵硬，是带有双叉钳子的工业操纵器，我们需要一些不同的东西。我个人认为我们正在越来越接近这个目标，因为我们开始重新想象机器人是什么。特别是，我们一直在研究软机器人和软机械手。我们已经证明，使用软机械手（我们可以在我的实验室中设计和制造的那种），我们能够比传统的两指抓握更可靠、更直观地拾取和处理物体。

Coming back to dexterity, on the hardware side, most of our robot hands still look like they did 50 years ago. Most of our robot hands are still very rigid, industrial manipulators with a two-pronged pincer, and we need something different. I personally believe that we are getting closer because we are beginning to look at reimagining what a robot is. In particular, we have been working on soft robots and soft robot hands. We’ve shown that with soft robot hands—the kind that we can design and build in my lab—we are able to pick up objects and handle objects much more reliably and much more intuitively than what is possible with traditional two-finger grasps.

它的工作原理如下：如果你有一个传统的机械手，其手指全部由金属制成，那么它们就可以进行技术上所谓的“硬手指接触”——你将手指放在试图抓住的物体上的某一点上，这就是你可以施加力和扭矩的点。如果你有这样的设置，那么你确实需要知道试图拿起的物体的精确几何形状。然后，你需要非常精确地计算将手指放在物体表面的位置，以便所有力和扭矩平衡，并且它们可以抵抗外力和扭矩。在技术文献中，这被称为“力闭合和形式闭合问题”。这个问题需要非常大量的计算、非常精确的执行以及对试图抓住的物体的非常准确的了解。

It works as follows: if you have a traditional robot hand where the fingers are all made out of metal, then they are capable of what is technically called “hard finger contact”—you put your finger on the object you’re trying to grasp at one point, and that’s the point at which you can exert forces and torques. If you have that kind of a setup, then you really need to know the precise geometry of the object that you’re trying to pick up. You then need to calculate very precisely where to put your fingers on the surface of the object so that all their forces and torques balance out, and they can resist external forces and torques. This is called in technical literature, “the force closure and form closure problem.” This problem requires very heavy computation, very precise execution, and very accurate knowledge of the object that you’re trying to grasp.

人类抓取物体时不会这样做。作为一个实验，试着用指甲抓住一个杯子——这是一项非常困难的任务。作为人类，你对物体及其位置有着完美的了解，但你会很难做到这一点。有了柔软的手指，你实际上不需要知道你试图抓住的物体的确切几何形状，因为手指会顺应物体的表面。沿着更宽的表面区域接触意味着你不必精确地考虑将手指放在哪里才能可靠地包裹和提起物体。

That’s not something that humans do when they grasp an object. As an experiment, try to grasp a cup with your fingernails—it is such a difficult task. As a human, you have a perfect knowledge of the object and where it is located, but you will have a difficult time with that. With soft fingers, you actually don’t need to know the exact geometry of the object you’re trying to grasp because the fingers will comply to whatever the object surface is. Contact along a wider surface area means that you don’t have to think precisely about where to place the fingers in order to reliably envelop and lift the object.

这意味着机器人将更加强大，算法也将更加简单。因此，我对未来抓握和操控方面的进步非常乐观。我认为软手以及软机器人将成为灵活性提升的一个关键方面，就像激光扫描仪是提升机器人导航能力的一个关键方面一样。

That translates into much more capable robots and much simpler algorithms. As a result, I’m very bullish about the future progress in grasping and manipulation. I think that soft hands, and in general, soft robots are going to be a very critical aspect of advancement in dexterity, just like the laser scanner was a critical aspect of advancing the navigation capabilities of robots.

这又回到了我的观察：机器是由身体和大脑组成的。如果你改变机器的身体，让它更强大，那么你就能够使用不同类型的算法来控制机器人。我对软机器人技术非常感兴趣，我对软机器人技术可能影响多年来停滞不前的机器人技术领域感到非常兴奋。在抓取和操纵方面已经取得了很大进展，但我们还没有与自然系统、人类或动物相媲美的能力。

That goes back to my observation that machines are made up of bodies and brains. If you change the body of the machine and you make it more capable, then you will be able to use different types of algorithms to control that robot. I’m very excited about soft robotics, and I’m very excited about the potential for soft robotics to impact an area of robotics that has been stagnant for many years. A lot of progress has been made in grasping and manipulation, but we do not have the kinds of capabilities that compare with those of natural systems, people or animals.

马丁·福特：我们来谈谈人工智能在向人类水平的人工智能（AGI）迈进方面取得的进展。这条道路是什么样的？我们离它还有多远？

MARTIN FORD: Let’s talk about progress in AI toward human-level artificial intelligence or AGI. What does that path look like, and how close are we?

DANIELA RUS：我们已经研究人工智能问题 60 多年了，如果该领域的创始人看到我们今天所吹捧的巨大进步，他们会非常失望，因为我们似乎并没有取得多大进展。我认为 AGI 对我们来说根本不是近期的未来。

DANIELA RUS: We have been working on AI problems for over 60 years, and if the founders of the field were able to see what we tout as great advances today, they would be very disappointed because it appears we have not made much progress. I don’t think that AGI is in the near future for us at all.

我认为大众媒体对人工智能是什么、什么不是人工智能存在很大误解。我认为今天，大多数人说“人工智能”，实际上是指机器学习，更重要的是，他们指的是机器学习中的深度学习。

I think that there is a great misunderstanding in the popular press about what artificial intelligence is and what it isn’t. I think that today, most people who say “AI,” actually mean machine learning, and more than that, they mean deep learning within machine learning.

我认为，如今谈论人工智能的大多数人都倾向于将这些术语的含义拟人化。非专家说“智能”这个词时，只会联想到一种东西，那就是人的智能。

I think that most people who talk about AI today tend to anthropomorphize what these terms mean. Someone who is not an expert says the word “intelligence” and only has one association with intelligence, and that is the intelligence of people.

当人们说“机器学习”时，他们会想象机器学习就像人类一样。然而，这些术语在技术背景下的含义截然不同。如果你想想机器学习如今能做什么，那绝对是非凡的。机器学习是一个从数百万个通常手动标记的数据点开始的过程，系统旨在学习数据中普遍存在的模式，或根据该数据做出预测。

When people say “machine learning,” they imagine that the machine learned just like a human has learned. Yet these terms mean such different things in the technical context. If you think about what machine learning can do today, it’s absolutely extraordinary. Machine learning is a process that starts with millions of usually manually labeled data points, and the system aims to learn a pattern that is prevalent in the data, or to make a prediction based on that data.

这些系统在这方面的表现比人类好得多，因为这些系统能够吸收和关联比人类更多的数据点。但是，当系统得知，例如，照片中有一个咖啡杯时，它实际上是在说，构成当前照片中代表咖啡杯的斑点的像素与人类在图像中标记为咖啡杯的其他斑点相同。系统并不知道那个咖啡杯代表什么。

These systems can do this much better than humans because these systems can assimilate and correlate many more data points then humans are able to. However, when a system learns, for example, that there is a coffee mug in a photograph, what it is actually doing is it’s saying that the pixels that form this blob that represents the coffee mug in the current photo are the same as other blobs that humans have labeled in images as coffee mugs. The system has no real idea what that coffee mug represents.

系统不知道如何处理它，它不知道你是喝了它、吃了它还是扔了它。如果我告诉你我的桌子上有一个咖啡杯，你不需要看到那个咖啡杯就能知道它是什么，因为你拥有当今机器所不具备的推理和经验。

The system has no idea what to do with it, it doesn’t know if you drink it, eat it, or if you throw it. If I told you that there is a coffee mug on my desk, you don’t need to see that coffee mug in order to know what it is because you have the kind of reasoning and experience that machines today simply do not have.

在我看来，这种智能与人类智能之间的差距是巨大的，我们需要很长时间才能达到这个水平。我们不知道定义我们自身智能的过程，也不知道我们的大脑是如何运作的。我们不知道孩子们是如何工作的。我们对大脑了解一点，但这些了解与需要了解的东西相比微不足道。对智能的理解是当今科学界最深刻的问题之一。我们看到神经科学、认知科学和计算机科学的交叉点正在取得进展。

To me, the gap between this and human-level intelligence is extraordinary, and it will take us a long time to get there. We have no idea of the processes that define our own intelligence, and no idea how our brain works. We have no idea how children work. We know a little bit about the brain, but that amount is insignificant to how much there is to know. The understanding of intelligence is one of the most profound questions in science today. We see progress at the intersection between neuroscience, cognitive science, and computer science.

马丁·福特：是否有可能出现非凡的突破，真正推动事情的发展？

MARTIN FORD: Is it possible that there might be an extraordinary breakthrough that really moves things along?

DANIELA RUS：这是有可能的。在我们的实验室里，我们非常有兴趣弄清楚我们是否可以制造出适应人类的机器人。我们开始研究我们是否可以检测和分类大脑活动，这是一个具有挑战性的问题。

DANIELA RUS: That’s possible. In our lab, we’re very interested in figuring out whether we can make robots that will adapt to people. We started looking at whether we can detect and classify brain activity, which is a challenging problem.

我们主要能够根据“你错了”信号（称为“错误相关电位”）判断一个人是否察觉到某事不对劲。每个人都会发出这种信号，与他们的母语和环境无关。借助我们今天拥有的外部传感器（称为脑电图帽），我们可以相当可靠地检测到“你错了”信号。这很有趣，因为如果我们能做到这一点，那么我们就可以想象出这样的应用：工人可以与机器人并肩工作，他们可以在远处观察机器人，并在检测到错误时纠正错误。事实上，我们有一个项目正在解决这个问题。

We are mostly able to classify whether a person detects that something is wrong because of the “you are wrong” signal—called the “error-related potential.” This is a signal that everyone makes, independent of their native tongue and independent of their circumstances. With the external sensors we have today, which are called EEG caps, we are fairly reliably able to detect the “you are wrong” signal. That’s interesting because if we can do that, then we can imagine applications where workers could work side by side with robots, and they could observe the robots from a distance and correct their mistakes when a mistake is detected. In fact, we have a project that addresses this question.

不过，有趣的是，这些脑电图帽由 48 个电极组成，放在你的头上——这是一个非常稀疏的机械装置，让你想起了计算机由杠杆组成的时代。另一方面，我们有能力进行侵入性手术，在神经细胞层面上利用神经元，所以你可以把探针插入人脑，这样你就可以非常精确地检测神经层面的活动。我们在外部能做的事情和我们在侵入性方面能做的事情之间存在很大差距，我想知道，在某个时候，我们是否会在感知大脑活动和以更高的分辨率观察脑电波活动方面取得某种摩尔定律式的进步。

What’s interesting, though, is that these EEG caps are made up of 48 electrodes placed on your head—it’s a very sparse, mechanical setup that reminds you of when computers were made up of levers. On the other hand, we have the ability to do invasive procedures to tap into neurons at the level of the neural cell, so you could actually stick probes into the human brain, and you could detect neural-level activity very precisely. There’s a big gap between what we can do externally and what we can do invasively, and I wonder whether at some point we will have some kind of Moore’s law improvement on sensing brain activity and observing brainwave activity at a much higher resolution.

马丁·福特：所有这些技术的风险和弊端是什么？一方面是对就业的潜在影响。我们是否正在面临一场可能消除大量工作的重大变革？这是我们必须考虑适应的事情吗？

MARTIN FORD: What about the risks and the downsides of all of this technology? One aspect is the potential impact on jobs. Are we looking at a big disruption that could eliminate a lot of work, and is that something we have to think about adapting to?

DANIELA RUS：当然！工作正在发生变化：工作正在消失，工作正在被创造。麦肯锡全球研究所发表了一项研究，提出了一些非常重要的观点。他们研究了许多职业，发现有些任务可以通过当今机器能力水平实现自动化，而其他任务则不能。

DANIELA RUS: Absolutely! Jobs are changing: jobs are going away, and jobs are being created. The McKinsey Global Institute published a study that gives some really important views. They looked at a number of professions and observed that there are certain tasks that can be automated with the level of machine capability today, and others that cannot.

如果你分析一下人们在各种职业中如何度过时间，就会发现工作有几种类型。人们花时间应用专业知识；与他人互动；管理；进行数据处理；进行数据输入；从事可预测的体力劳动；从事不可预测的体力劳动。归根结底，有些任务可以自动化，有些则不能。可预测的体力劳动和数据任务是可以利用当今技术实现自动化的常规任务，但其他任务则不能。

If you do an analysis of how people spend time in various professions, there are certain categories of work. People spend time applying expertise; interacting with others; managing; doing data processing; doing data entry; doing predictable physical work; doing unpredictable physical work. Ultimately, there are tasks that can be automated and tasks that can’t. The predictable physical work and the data tasks are routine tasks that can be automated with today’s technologies, but the other tasks can’t.

我对此非常受启发，因为我看到技术可以减轻我们的日常工作负担，让我们有时间专注于工作中更有趣的部分。让我们来看一个医疗保健领域的例子。我们有一辆自动轮椅，我们一直在与理疗师讨论如何使用这辆轮椅。他们对此非常兴奋，因为目前，理疗师在医院里以以下方式与患者合作：

I’m actually very inspired by this because what I see is that technology can relieve us of routine work in order to give us time to focus on the more interesting parts of our work. Let’s go through an example in healthcare. We have an autonomous wheelchair, and we have been talking with physical therapists about using this wheelchair. They are very excited about it because, at the moment, the physical therapist works with patients in the hospital in the following way:

对于每个新病人，理疗师必须去病人的病床，把病人放在轮椅上，推到健身房，然后一起在健身房锻炼，一小时后，理疗师必须把病人送回病人的病床。大量的时间都花在了移动病人上，而不是病人护理上。

For every new patient, the physical therapist has to go to the patient’s bed where they have to put the patient in a wheelchair, push the patient to the gym where they’ll work together in the gym and at the end of the hour, the physical therapist has to take the patient back to the patient’s hospital bed. A significant amount of time is spent moving the patient around and not on patient care.

现在想象一下，如果理疗师不必这样做会怎么样。想象一下，如果理疗师可以留在健身房，而病人会坐着自动轮椅出现。那么病人和理疗师都会有更好的体验。病人会从理疗师那里得到更多的帮助，而理疗师会专注于运用他们的专业知识。我对提高我们在工作中度过的时间质量和提高工作效率的可能性感到非常兴奋。

Now imagine if the physical therapist didn’t have to do this. Imagine if the physical therapist could stay in the gym, and the patient would show up delivered by an autonomous wheelchair. Then both the patient and the physical therapist would have a much better experience. The patient would get more help from the physical therapist, and the physical therapist would focus on applying their expertise. I’m very excited about the possibility of enhancing the quality of time that we spend in our jobs and increasing our efficiency in our jobs.

第二个观察结果是，一般来说，我们更容易分析哪些行业可能会消失，而不是想象哪些行业可能会回归。例如，在 20 世纪，美国的农业就业率从 40% 下降到 2%。20 世纪没有人猜到会发生这种情况。想想看，就在 10 年前，当计算机行业蓬勃发展时，没有人预测到社交媒体、应用商店、云计算，甚至大学咨询等其他行业的就业水平。如今有很多工作岗位在 10 年前并不存在，人们也没有预料到它们会存在，它们雇用了很多人。我认为，思考未来的可能性以及技术将创造的新工作类型是令人兴奋的。

A second observation is that in general, it is much easier for us to analyze what might go away than to imagine what might come back. For instance, in the 20th century, agricultural employment dropped from 40% to 2% in the United States. Nobody in the 20th century guessed that this would happen. Just consider, then, that only 10 years ago, when the computer industry was booming, nobody predicted the level of employment in social media; in app stores; in cloud computing; and even in other things like college counseling. There are so many jobs that employ a lot of people today that did not exist 10 years ago, and that people did not anticipate would exist. I think that it’s exciting to think about the possibilities for the future and the new kinds of jobs that will be created as a result of technology.

马丁·福特：那么，您认为技术所摧毁的就业岗位和新创造的就业岗位会保持平衡吗？

MARTIN FORD: So, you think the jobs destroyed by technology and the new jobs created will balance out?

DANIELA RUS：嗯，我也有顾虑。一个顾虑是工作质量。有时，当你引进技术时，技术会让竞争更加公平。例如，过去出租车司机必须具备很多专业知识——他们必须有很强的空间推理能力，还必须记住大地图。随着 GPS 的出现，这种技能水平不再需要。这为更多人加入驾驶市场打开了空间，这往往会降低工资。

DANIELA RUS: Well, I do also have concerns. One concern is in the quality of jobs. Sometimes, when you introduce technology, the technology levels the playing field. For instance, it used to be that taxi drivers had to have a lot of expertise—they had to have great spatial reasoning, and they had to memorize large maps. With the advent of GPS, that level of skill is no longer needed. What that does is open the field for many more people to join the driving market, and that tends to lower the wages.

另一个担忧是，人们是否能接受足够的培训，以胜任技术带来的好工作。我认为只有两种方法可以应对这一挑战。在短期内，我们必须想办法帮助人们重新培训自己，如何帮助人们获得完成现有工作所需的技能。我每天都会听到无数次这样的声音：“我们需要你的人工智能学生。你能给我们派一些人工智能学生吗？”每个人都想要人工智能和机器学习方面的专家，所以有很多工作机会，也有很多人在找工作。然而，需求的技能不一定是人们拥有的技能，所以我们需要再培训计划来帮助人们获得这些技能。

Another concern is that I wonder if people are going to be trained well enough for the good jobs that will be created as a result of technology. I think that there are only two ways to approach this challenge. In the short term, we have to figure out how to help people retrain themselves, how to help people gain the skills that are needed in order to fulfill some of the jobs that exist. I can’t tell you how many times a day I hear, “We want your AI students. Can you send us any AI students?” Everyone wants experts in artificial intelligence and machine learning, so there are a lot of jobs, and there are also a lot of people who are looking for jobs. However, the skills that are in demand are not necessarily the skills that people have, so we need retraining programs to help people acquire those skills.

我坚信任何人都可以学习技术。我最喜欢的例子是一家名为 BitSource 的公司。BitSource 几年前在肯塔基州成立，该公司正在将煤矿工人重新培训为数据挖掘人员，并取得了巨大的成功。该公司培训了许多失业的矿工，现在他们能够找到更好、更安全、更愉快的工作。这个例子实际上告诉我们，通过正确的计划和正确的支持，我们实际上可以帮助人们度过这个过渡时期。

I’m a big believer in the fact that actually anybody can learn technology. My favorite example is a company called BitSource. BitSource was launched a couple of years back in Kentucky, and this company is retraining coal miners into data miners and has been a huge success. This company has trained a lot of the miners who lost their jobs and who are now in a position to get much better, much safer and much more enjoyable jobs. It’s an example that actually tells us that with the right programs and the right support, we can actually help people in this transition period.

马丁·福特：这仅仅是就工人的再培训而言，还是我们需要从根本上改变我们的整个教育体系？

MARTIN FORD: Is that just in terms of retraining workers, or do we need to fundamentally change our entire educational system?

DANIELA RUS：20 世纪，我们通过阅读、写作和算术来定义读写能力。在 21 世纪，我们应该扩展读写能力的含义，并加入计算思维。如果我们在学校教授如何制作东西以及如何通过编程为它们注入生命，我们将赋予学生力量。我们可以让他们达到可以想象任何事物并将其变为现实的程度，并且他们将拥有实现这些事物的工具。更重要的是，到他们完成高中教育时，这些学生将具备未来所需的技术技能，他们将接触到一种不同的学习方式，使他们能够为未来提供帮助。

DANIELA RUS: In the 20th century we had reading, writing, and arithmetic that defined literacy. In the 21st century, we should expand what literacy means, and we should add computational thinking. If we teach in schools how to make things and how to breathe life into them by programming, we will empower our students. We can get them to the point where they can imagine anything and make it happen, and they will have the tools to make it happen. More importantly, by the time they finish high school, these students will have the technical skills that will be required in the future, and they will be exposed to a different way of learning that will enable them to help themselves for the future.

关于未来工作，我最后想说的是，我们对学习的态度也必须改变。今天，我们采用的是学习和工作的顺序模式。我的意思是，大多数人一生中都会花一部分时间学习，到了某个时候，他们会说：“好了，我们学完了，现在要开始工作了。”然而，随着技术的发展和新能力的出现，我认为重新考虑学习的顺序方法非常重要。我们应该考虑一种更加并行的学习和工作方式，我们将乐于获取新技能，并将这些技能作为终身学习的过程。

The final thing I want to say about the future of work is that our attitude toward learning will also have to change. Today, we operate with a sequential model of learning and working. What I mean by this is that most people spend some chunk of their lives studying and at some point, they say, “OK, we’re done studying, now we’re going to start working.” With technology accelerating and bringing in new types of capabilities, though, I think it’s very important to reconsider the sequential approach to learning. We should consider a more parallel approach to learning and working, where we will be open to acquiring new skills and applying those skills as a lifelong learning process.

马丁·福特：一些国家将人工智能作为战略重点，或制定了针对人工智能和机器人的明确产业政策。尤其是中国，正在大力投资这一领域。您是否认为，在先进人工智能领域存在竞争，而美国是否有落后的风险？

MARTIN FORD: Some countries are making AI a strategic focus or adopting an explicit industrial policy geared toward AI and robotics. China, in particular, is investing massively in this area. Do you think that there is a race toward advanced AI, and is the US at risk of falling behind?

DANIELA RUS：当我看到世界各地人工智能的发展时，我觉得这真是太神奇了。中国、加拿大、法国、英国等几十个国家都在大力投资人工智能。许多国家都把未来押注在人工智能上，我认为美国也应该这样做。我认为我们应该考虑人工智能的潜力，我们应该增加对人工智能的支持和资助。

DANIELA RUS: When I look at what is happening in AI around the world, I think it is amazing. You have China, Canada, France, and the UK, among dozens of others, hugely investing in AI. Many countries are betting their future on AI, and I think we in the US should do too. I think we should consider the potential for AI, and we should increase the support and the funding of AI.

DANIELA RUS 是麻省理工学院电气工程和计算机科学系的 Andrew (1956) 和 Erna Viterbi 教授，也是计算机科学和人工智能实验室 (CSAIL) 主任。Daniela 的研究兴趣是机器人技术、人工智能和数据科学。

DANIELA RUS is the Andrew (1956) and Erna Viterbi Professor of Electrical Engineering and Computer Science and Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT. Daniela’s research interests are in robotics, artificial intelligence, and data science.

她的工作重点是开发自主科学和工程，以实现长期目标，即在未来实现机器广泛融入生活结构，帮助人们完成认知和体力任务。她的研究解决了机器人现状与机器人普及前景之间的一些差距：提高机器在以人为中心的环境中推理、学习和适应复杂任务的能力，开发机器人与人之间的直观界面，以及创建快速高效设计和制造新机器人的工具。这项工作的应用范围很广，包括交通、制造、农业、建筑、环境监测、水下探索、智能城市、医学以及烹饪等家庭任务。

The focus of her work is developing the science and engineering of autonomy, toward the long-term objective of enabling a future with machines pervasively integrated into the fabric of life, supporting people with cognitive and physical tasks. Her research addresses some of the gaps between where robots are today and the promise of pervasive robots: increasing the ability of machines to reason, learn, and adapt to complex tasks in human-centered environments, developing intuitive interfaces between robots and people, and creating the tools for designing and fabricating new robots quickly and efficiently. The applications of this work are broad and include transportation, manufacturing, agriculture, construction, monitoring the environment, underwater exploration, smart cities, medicine, and in-home tasks such as cooking.

Daniela 担任麻省理工学院智能核心探索中心副主任，以及丰田-CSAIL 联合研究中心主任，该中心的重点是推动人工智能研究及其在智能汽车中的应用。她是丰田研究所顾问委员会成员。

Daniela serves as the Associate Director of MIT’s Quest for Intelligence Core, and as Director of the Toyota-CSAIL Joint Research Center, whose focus is the advancement of AI research and its applications to intelligent vehicles. She is a member of the Toyota Research Institute advisory board.

Daniela 是 2002 届麦克阿瑟奖获得者、ACM、AAAI 和 IEEE 研究员，也是美国国家工程院和美国艺术与科学学院院士。她是机器人工业协会颁发的 2017 年 Engelberger 机器人奖的获得者。她在康奈尔大学获得计算机科学博士学位。

Daniela is a Class of 2002 MacArthur Fellow, a fellow of ACM, AAAI and IEEE, and a member of the National Academy of Engineering and the American Academy of Arts and Sciences. She is the recipient of the 2017 Engelberger Robotics Award from the Robotics Industries Association. She earned her PhD in Computer Science from Cornell University.

丹妮拉还与 Pilobolus 舞蹈公司合作开展了两个融合科技与艺术的项目。《Seraph》是一部讲述人机友谊的田园故事，编排于 2010 年，并于 2010-2011 年在波士顿和纽约市演出。《Umbrella Project》是一部探索群体行为的参与式表演，编排于 2012 年，并在剑桥、巴尔的摩和新加坡的 PopTech 2012 上演出。

Daniela has also worked on two collaborative projects with the Pilobolus Dance company at the intersection of technology and art. Seraph, a pastoral story about human-machine friendship, was choreographed in 2010 and performed in 2010-2011 in Boston and New York City. The Umbrella Project, a participatory performance exploring group behavior, was choreographed in 2012 and performed at PopTech 2012, in Cambridge, Baltimore, and Singapore.

麦肯锡全球研究院董事长兼院长

CHAIRMAN AND DIRECTOR OF MCKINSEY GLOBAL INSTITUTE

詹姆斯是麦肯锡的高级合伙人兼麦肯锡全球研究所主席，研究全球经济和技术趋势。詹姆斯为许多世界领先科技公司的首席执行官和创始人提供咨询。他领导人工智能和数字技术及其对组织、工作和全球经济影响的研究。詹姆斯被奥巴马总统任命为白宫全球发展委员会副主席，并被美国商务部长任命为数字经济委员会和国家创新委员会副主席。他是牛津互联网研究所、麻省理工学院数字经济计划、斯坦福大学人工智能百年研究的董事会成员，也是 DeepMind 的研究员。

James is a senior partner at McKinsey and Chairman of the McKinsey Global Institute, researching global economic and technology trends. James consults with the chief executives and founders of many of the world’s leading technology companies. He leads research on AI and digital technologies and their impact on organizations, work, and the global economy. James was appointed by President Obama as vice chair of the Global Development Council at the White House and by US Commerce Secretaries to the Digital Economy Board and National Innovation Board. He is on the boards of the Oxford Internet Institute, MIT’s Initiative on the Digital Economy, the Stanford-based 100-Year Study on AI, and he is a fellow at DeepMind.

马丁·福特：我想我们可以先回顾一下您的学术和职业轨迹。我知道您来自津巴布韦。您是如何对机器人和人工智能产生兴趣，然后最终在麦肯锡担任现在的职务的？

MARTIN FORD: I thought we could start by having you trace your academic and career trajectory. I know you came from Zimbabwe. How did you get interested in robotics and artificial intelligence and then end up in your current role at McKinsey?

詹姆斯·马尼卡：我生长在一个黑人隔离的小镇，那里当时叫罗得西亚，后来成为津巴布韦。我一直受到科学理念的启发，部分原因是我的父亲是 20 世纪 60 年代初第一位从津巴布韦来到美国的黑人富布赖特学者。在那里，我父亲参观了卡纳维拉尔角的 NASA，在那里他看到了火箭飞上天空。在我很小的时候，他从美国回来后，父亲就给我灌输了科学、太空和技术的理念。所以我在这个种族隔离的小镇长大，思考着科学和太空，用我能找到的任何东西制作模型飞机和机器。

JAMES MANYIKA: I grew up in a segregated black township in what was then Rhodesia, before it became Zimbabwe. I was always inspired by the idea of science, partly because my father had been the first black Fulbright scholar from Zimbabwe to come to the United States of America in the early 1960s. While there, my father visited NASA at Cape Canaveral, where he watched rockets soar up into the sky. And in my early childhood after he came back from America, my father filled my head with the idea of science, space, and technology. So, I grew up in this segregated township, thinking about science and space, building model planes and machines out of whatever I could find.

当我在该国成为津巴布韦后上大学时，我的本科学位是电气工程，大量涉及数学和计算机科学。在那里，多伦多大学的一位客座研究员让我参与了一个神经网络项目。那时我了解了鲁梅尔哈特反向传播和 Logisti S 函数在神经网络算法中的应用。

When I got to university after the country had become Zimbabwe, my undergraduate degree was in electrical engineering with heavy doses of mathematics and computer science. And while there a visiting researcher from the University of Toronto got me involved in a project on neural networks. That’s when I learned about Rumelhart Backpropagation and the use of logisti sigmoid functions in neural network algorithms.

很快，我凭借优异的成绩获得了罗德奖学金，进入牛津大学学习。我进入编程研究小组，在托尼·霍尔手下工作。霍尔因发明快速排序和痴迷于形式化方法和编程语言的公理规范而闻名。我攻读数学和计算机科学硕士学位，在数学证明以及算法的开发和验证方面投入了大量精力。那时，我已经放弃了当宇航员的想法，但我想，至少如果我从事机器人和人工智能方面的工作，我可能会接近与太空探索相关的科学。

Fast forward, I did well enough to get a Rhodes scholarship to go to Oxford University, where I was in the Programming Research Group, working under Tony Hoare, who is best known for inventing Quicksort and for his obsession with formal methods and axiomatic specifications of programming languages. I studied for a master’s degree in mathematics and computer science and worked a lot on mathematical proofs and the development and verification of algorithms. By this time, I’d given up on the idea that I would be an astronaut, but I thought that at least if I worked on robotics and AI, I might get close to science related to space exploration.

我最终进入了牛津大学的机器人研究小组，他们当时正在研究人工智能，但当时并没有多少人这么称呼它，因为当时人工智能有负面含义，因为当时刚刚经历了一场“人工智能寒冬”，或者说一系列寒冬，人工智能没有达到人们的预期和期望。所以他们把自己的工作称为除了人工智能以外的所有东西——机器感知、机器学习、机器人技术或简单的神经网络；但当时没有人愿意把自己的工作称为人工智能。现在我们遇到了相反的问题，每个人都想把所有东西都称为人工智能。

I wound up in the Robotics Research Group at Oxford, where they were actually working on AI, but not many people called it that in those days because AI had a negative connotation at the time, after what had recently been a kind of “AI Winter” or a series of winters, where AI had underdelivered on its hype and expectations. So, they called their work everything but AI—it was machine perception, machine learning, it was robotics or just plain neural networks; but no-one in those days was comfortable calling their work AI. Now we have the opposite problem, everyone wants to call everything AI.

马丁·福特：这是什么时候的事？

MARTIN FORD: When was this?

詹姆斯·马尼卡：那是 1991 年，当时我在牛津大学机器人研究组开始攻读博士学位。这段职业生涯让我有机会与机器人和人工智能领域的许多人一起工作。因此，我认识了安德鲁·布莱克和莱昂内尔·塔拉森科等人，他们当时正在研究神经网络；迈克尔·布雷迪（现为迈克尔爵士）当时正在研究机器视觉；我还认识了休·杜兰特-怀特，当时正在研究分布式智能和机器人系统。他成为了我的博士导师。我们一起制造了几辆自动驾驶汽车，还一起写了一本书，书中借鉴了我们正在开发的研究和智能系统。

JAMES MANYIKA: This was in 1991, when I started my PhD at the Robotics Research Group at Oxford. This part of my career really opened me to working with a number of different people in the robotics and AI fields. So, I met people like Andrew Blake and Lionel Tarassenko, who were working on neural networks; Michael Brady, now Sir Michael, who was working on machine vision; and I met Hugh Durrant-Whyte, who was working on distributed intelligence and robotic systems. He became my PhD advisor. We built a few autonomous vehicles together and we also wrote a book together drawing on the research and intelligence systems we were developing.

通过我所做的研究，我最终与 NASA 喷气推进实验室的一个团队合作，该团队正在研究火星探测器。NASA 有兴趣将他们正在开发的机器感知系统和算法应用于火星探测器项目。我想这是我最接近太空的一次了！

Through the research I was doing, I wound up collaborating with a team at the NASA Jet Propulsion Laboratory that was working on the Mars rover vehicle. NASA was interested in applying the machine perception systems and algorithms that they were developing to the Mars rover vehicle project. I figured that this was as close as I’m ever going to get to go into space!

马丁·福特：那么，你编写的某些代码确实在火星探测器上运行了？

MARTIN FORD: So, there was actually some code that you wrote running on the rover, on Mars?

詹姆斯·马尼卡：是的，我当时在加利福尼亚州帕萨迪纳市喷气推进实验室的人机系统小组工作。我是那里几位客座科学家之一，研究这些机器感知和导航算法，其中一些算法被用在模块化和自动驾驶汽车系统以及其他地方。

JAMES MANYIKA: Yes, I was working with the Man Machine Systems group at JPL in Pasadena, California. I was one of several visiting scientists there working on these machine perception and navigation algorithms, and some of them found their way onto the modular and autonomous vehicle systems and other places.

我在牛津大学机器人研究小组的那段时间真正激发了我对人工智能的兴趣。我发现机器感知特别令人着迷：如何为分布式和多智能体系统构建学习算法，如何使用机器学习算法来理解环境，以及如何开发能够自主构建这些环境模型的算法，尤其是那些你之前没有了解过的环境，必须边走边学——比如火星表面。

That period at Oxford in the Robotics Research Group is what really sparked my interest in AI. I found machine perception particularly fascinating: the challenges of how to build learning algorithms for distributed and multi-agent systems, how to use machine learning algorithms to make sense of environments, and how to develop algorithms that could autonomously build models of those environments, in particular, environments where you had no prior knowledge of them and had to learn as you go—like the surface of Mars.

我从事的很多工作不仅应用于机器视觉，还应用于分布式网络、传感和传感器融合。我们构建了这些基于神经网络的算法，这些算法结合了 Judea Pearl 开创的贝叶斯网络、卡尔曼滤波器和其他估计和预测算法，本质上是构建机器学习系统。我们的想法是，这些系统可以从环境中学习，从来自各种不同质量来源的输入数据中学习，并做出预测。它们可以绘制地图并收集所处环境的知识；然后它们可能能够做出预测和决策，就像智能​​系统一样。

A lot of what I was working on had applications not just in machine vision, but in distributed networks and sensing and sensor fusion. We were building these neural network-based algorithms that were using a combination of Bayesian networks of the kind Judea Pearl had pioneered, Kalman filters and other estimation and prediction algorithms to essentially build machine learning systems. The idea was that these systems could learn from the environment, learn from input data from a wide range of sources of varying quality, and make predictions. They could build maps and gather knowledge of the environments that they were in; and then they might be able to make predictions and decisions, much like intelligent systems would.

因此，我最终在麻省理工学院担任客座教授期间结识了罗德尼·布鲁克斯，现在我们仍然是朋友。当时，我与他仍是好友。当时，我与麻省理工学院的机器人小组和海洋资助项目合作，共同开发水下机器人。在此期间，我还结识了伯克利大学机器人和人工智能教授斯图尔特·罗素等人，因为他曾在牛津大学我的研究小组工作过。事实上，我当时的许多同事都继续从事开创性的工作，比如约翰·伦纳德（现为麻省理工学院机器人学教授）和 DeepMind 的安德鲁·齐瑟曼。尽管我已涉足商业和经济学的其他领域，但我一直密切关注人工智能和机器学习领域的工作，并尽我所能跟上潮流。

So, eventually I met Rodney Brooks, who I’m still friends with today, during my visiting faculty fellowship at MIT, where I was working with the Robotics Group at MIT and the Sea Grant project that was building underwater robots. During this time, I also got to know people like Stuart Russell, who’s a professor in robotics and AI at Berkeley, because he had spent time at Oxford in my research group. In fact, many of my colleagues from those days have continued to do pioneering work, people like John Leonard, now a Robotics Professor at MIT and Andrew Zisserman, at DeepMind. Despite the fact that I’ve wandered off into other areas in business and economics, I’ve stayed close to the work going on in AI and machine learning and try to keep up as best as I can.

马丁·福特：那么，考虑到您在牛津大学任教，您一开始就非常注重技术性吗？

MARTIN FORD: So, you started out with a very technical orientation, given that you were teaching at Oxford?

詹姆斯·曼尼卡：是的，我是牛津大学贝利奥尔学院的教员和研究员，我给学生讲授数学和计算机科学课程，同时还参与我们在机器人技术方面的一些研究。

JAMES MANYIKA: Yes, I was on the faculty and a fellow at Balliol College, Oxford, and I was teaching students courses in mathematics and computer science and as well as on some of the research we were doing in robotics.

马丁·福特：从那里跳槽到麦肯锡的商业和管理咨询部门，听起来相当不寻常。

MARTIN FORD: It sounds like a pretty unusual jump from there to business and management consulting at McKinsey.

詹姆斯·马尼卡：这其实是偶然的。我最近订婚了，还收到了麦肯锡的邀请，让我加入硅谷；我认为去麦肯锡工作是一次短暂而有趣的旅程。

JAMES MANYIKA: That was actually as much by accident as anything else. I’d recently become engaged, and I had also received an offer from McKinsey to join them in Silicon Valley; and I thought it would be a brief, interesting detour, to go to McKinsey.

当时，我和许多朋友、同事（比如和我一起在机器人研究实验室工作的 Bobby Rao）一样，对开发能够参加 DARPA 无人驾驶汽车挑战赛的系统很感兴趣。这是因为我们的很多算法都适用于自动驾驶汽车和无人驾驶汽车，而当时，DARPA 挑战赛是可以应用这些算法的地方之一。当时我所有的朋友都搬到了硅谷。Bobby 当时在伯克利做博士后，与 Stuart Russell 等人一起工作，所以我认为我应该接受麦肯锡在旧金山的邀请。这是一种接近硅谷的方式，可以接近包括 DARPA 挑战赛在内的一些活动的举办地。

At the time, like many of my friends and colleagues, such as Bobby Rao, who had also been in the Robotics Research Lab with me, I was interested in building systems that could compete in the DARPA driverless car challenge. This was because a lot of our algorithms were applicable to autonomous vehicles and driverless cars and back then, the DARPA challenge was one of the places where you could apply those algorithms. All of my friends were moving to Silicon Valley then. Bobby was at that time a post-doc at Berkeley working with Stuart Russell and others, and so I thought I should take this McKinsey offer in San Francisco. It was a way to be close to Silicon Valley and to be close to where some of the action, including the DARPA challenge, was taking place.

马丁·福特：您现在在麦肯锡担任什么职务？

MARTIN FORD: What is your role now at McKinsey?

詹姆斯·马尼卡：我最终做了两件事。一是与硅谷的许多先锋科技公司合作，我有幸与许多创始人和首席执行官共事，并为他们提供建议。另一部分工作随着时间的推移而不断发展，即领导技术交叉领域及其对商业和经济影响的研究。我是麦肯锡全球研究所的主席，我们不仅研究技术，还研究宏观经济和全球趋势，以了解它们对商业和经济的影响。我们很荣幸能拥有出色的学术顾问，其中包括对技术影响进行深入思考的经济学家，如埃里克·布林约尔夫森、哈尔·瓦里安和诺贝尔奖获得者迈克·斯宾塞，甚至过去的鲍勃·索洛。

JAMES MANYIKA: I’ve ended up doing two kinds of things. One is working with many of the pioneering technology companies in Silicon Valley, where I have been fortunate to work with and advise many founders and CEOs. The other part, which has grown over time, is leading research at the intersection of technology and its impact on business and the economy. I’m the chairman of the McKinsey Global Institute, where we research not just technology but also macroeconomic and global trends to understand their impact on business and the economy. We are privileged to have amazing academic advisors that include economists who also think a lot about technology’s impacts, people like Erik Brynjolffson, Hal Varian, and Mike Spence, the Nobel laureate, and even Bob Solow in the past.

将其与人工智能联系起来，我们一直在关注颠覆性技术，并跟踪人工智能的发展，我一直与 Eric Horvitz、Jeff Dean、Demis Hassabis 和李飞飞等人工智能朋友保持对话和合作，也向 Barbara Grosz 等传奇人物学习。虽然我试图贴近技术和科学，但我和麦肯锡全球研究院的同事花了更多时间思考和研究这些技术的经济和商业影响。

To link this back to AI, we’ve been looking a lot at disruptive technologies, and tracking the progress of AI, and I’ve stayed in constant dialogues as well as collaborated with AI friends like Eric Horvitz, Jeff Dean, Demis Hassabis, and Fei-Fei Li, and also learning from legends like Barbara Grosz. While I’ve tried to stay close to the technology and the science, my MGI colleagues and I have spent more time thinking about and researching the economic and business impacts of these technologies.

马丁·福特：我当然想深入探讨其对经济和就业市场的影响，但让我们先谈谈人工智能技术。

MARTIN FORD: I definitely want to delve into the economic and job market impact, but let’s begin by talking about AI technology.

您曾提到，早在 20 世纪 90 年代，您就开始研究神经网络。过去几年，深度学习出现了爆炸式增长。您对此有何感想？您是否认为深度学习是未来的圣杯，还是被过度炒作了？

You mentioned that you were working on neural networks way back in the 1990s. Over the past few years, there’s been this explosion in deep learning. How do you feel about that? Do you see deep learning as the holy grail going forward, or has it been overhyped?

詹姆斯·马尼卡：我们才刚刚发现深度学习和神经网络等技术及其多种形式的威力，以及强化学习和迁移学习等其他技术。这些技术都还有巨大的发展空间；我们只是触及了它们能带给我们的深层意义。

JAMES MANYIKA: We’re only just discovering the power of techniques such as deep learning and neural networks in their many forms, as well as other techniques like reinforcement learning and transfer learning. These techniques all still have enormous headroom; we’re only just scratching the surface of where they can take us.

深度学习技术正在帮助我们解决大量特定问题，无论是图像和对象分类、自然语言处理还是生成式人工智能（我们可以预测和创建序列和输出，无论是语音、图像等）。我们将在有时被称为“狭义人工智能”的领域取得巨大进展，即使用这些深度学习技术解决特定领域和问题。

Deep learning techniques are helping us solve a huge number of particular problems, whether it’s in image and object classification, natural language processing or generative AI, where we predict and create sequences and outputs whether its speech, images, and so forth. We’re going to make a lot of progress in what is sometimes called “narrow AI,” that is, solving particular areas and problems using these deep learning techniques.

相比之下，我们在有时被称为“通用人工智能”或 AGI 的领域进展较慢。虽然我们最近取得的进展比很长一段时间以来都要多，但我仍然认为 AGI 的进展会慢得多，因为它涉及一组更复杂、更困难的问题需要回答，需要更多的突破。

In comparison, we’re making slower progress on what is sometimes called “artificial general intelligence” or AGI. While we’ve made more progress recently than we’ve done in a long time, I still think progress is going to be much, much slower towards AGI, just because it involves a much more complex and difficult set of questions to answer and will require many more breakthroughs.

我们需要弄清楚如何思考迁移学习之类的问题，因为人类做得非常好的事情之一就是能够在这里学习一些东西，然后能够将这种学习应用到全新的环境中或之前从未遇到过的问题中。肯定会有一些令人兴奋的新技术出现，无论是在强化学习还是模拟学习中——AlphaZero 已经开始做的事情——你可以自我学习和自我创建结构，并开始解决更广泛和不同类型的挑战，就 AlphaZero 而言，不同类型的游戏。在另一个方向上，Jeff Dean 和 Google Brain 的其他人正在使用 AutoML 进行的工作确实令人兴奋。从帮助我们开始在自我设计的机器和神经网络方面取得进展的角度来看，这非常有趣。这些只是几个例子。可以说所有这些进展都在推动我们走向 AGI。但这些真的只是小步骤；还需要做很多很多，还有整个高级推理领域等等，我们几乎不知道如何解决。这就是我认为 AGI 还有很长的路要走的原因。

We need to figure out how to think about problems like transfer learning, because one of the things that humans do extraordinarily well is being able to learn something, over here, and then to be able to apply that learning in totally new environments or on a previously unencountered problem, over there. There are definitely some exciting new techniques coming up, whether in reinforcement learning or even simulated learning—the kinds of things that AlphaZero has begun to do—where you self-learn and self-create structures, as well start to solve wider and different categories of challenges, in the case of AlphaZero different kinds of games. In another direction the work that Jeff Dean and others at Google Brain are doing using AutoML is really exciting. That’s very interesting from the point of helping us start to make progress in machines and neural networks that design themselves. These are just a few examples. One could say all of this progress is nudging us towards AGI. But these are really just small steps; much, much more is needed, there are whole areas of high-level reasoning etc. that we barely know how to tackle. This is why I think AGI is still quite a long way away.

深度学习肯定会帮助我们实现狭义的人工智能应用。我们将看到大量应用已经转化为新产品和新公司。同时，值得指出的是，机器学习的使用和应用仍然存在一些实际限制，我们已经在一些 MGI 工作中指出了这一点。

Deep learning is certainly going to help us with narrow AI applications. We’re going to see lots and lots and lots and lots of applications that are already being turned into new products and new companies. At the same time, it’s worth pointing out that there are still some practical limitations to the use and application of machine learning, and we have pointed this out in some of our MGI work.

马丁·福特：您能举个例子吗？

MARTIN FORD: Do you have any examples?

JAMES MANYIKA：例如，我们知道许多此类技术仍然很大程度上依赖于标记数据，而在标记数据的可用性方面仍然存在很多限制。这通常意味着人类必须标记基础数据，这可能是一项庞大且容易出错的工作。事实上，一些自动驾驶汽车公司正在雇用数百人来手动注释来自原型汽车的数小时视频，以帮助训练算法。现在出现了一些新技术来解决需要标记数据的问题，例如由 Eric Horvitz 等人首创的流内监督；使用生成对抗网络或 GAN 等技术，这是一种半监督技术，通过这种技术可以生成可用数据，从而减少对需要人类标记的数据集的需求。

JAMES MANYIKA: For example, we know that many of these techniques still largely rely on labelled data, and there’s still lots of limitations in terms of the availability of labelled data. Often this means that humans must label underlying data, which can be a sizable and error-prone chore. In fact, some autonomous vehicle companies are hiring hundreds of people to manually annotate hours of video from prototype vehicles to help train the algorithms. There are some new techniques that are emerging to get around the issue of needing labeled data, for example, in-stream supervision pioneered by Eric Horvitz and others; the use of techniques like Generative Adversarial Networks or GANs, which is a semi-supervised technique through which usable data can be generated in a way that reduces the need for datasets that require labeling by humans.

但是我们仍然面临第二个挑战，那就是需要如此庞大而丰富的数据集。有趣的是，你可以通过观察哪些领域可以获得大量可用数据，或多或少地识别出那些正在取得惊人进展的领域。因此，我们在机器视觉方面取得的进展比其他应用更大也就不足为奇了，因为每天都有大量的图像和视频被上传到互联网上。现在，有一些很好的理由——监管、隐私、安全等——可能会在一定程度上限制数据的可用性。这也可以部分解释为什么不同的社会在提供数据方面会经历不同的进展速度。人口众多的国家自然会产生更多的数据，而不同的数据使用标准可能会使访问大型健康数据集（例如）并用它来训练算法变得更容易。因此，在中国，鉴于可用的数据集更大，你可能会看到人工智能在基因组学和“组学”方面的应用取得更大的进展。

But then we still have a second challenge of needing such large and rich data sets. It is quite interesting that you can more or less identify those areas that are making spectacular progress simply by observing which areas have access to a huge amount of available data. So, it is no surprise that we have made more progress in machine vision than in other applications, because of the huge volume of images and now video being put on the internet every day. Now, there are some good reasons—regulatory, privacy, security, and otherwise—that may limit data availability to some extent. And this can also, in part, explain why different societies are going to experience differential rates of progress on making data available. Countries with large populations, naturally, generate larger volumes of data, and different data use standards may make it easier to access large health data sets, for example, and use that to train algorithms. So, in China you might see more progress in using AI in genomics and “omics” given larger available data sets.

因此，数据可用性非常重要，这或许可以解释为什么某些领域的人工智能应用在某些地方比其他地方发展得更快。但我们也面临其他限制，比如我们仍然没有通用的人工智能工具，我们仍然不知道如何解决人工智能中的一般问题。事实上，有趣的是，你可能已经看到了，人们现在开始定义图灵测试的新形式。

So, data availability is a big deal and may explain why some areas of AI applications take off much faster in some places than others. But we’ve also got other limitations to deal with, like we still don’t have generalized tools in AI and we still don’t know how to solve general problems in AI. In fact, one of the fun things, and you may have seen this, is that people are now starting to define new forms of what used to be the Turing test.

马丁·福特：新的图灵测试？它将如何运作？

MARTIN FORD: A new Turing Test? How would that work?

詹姆斯·马尼卡：苹果联合创始人史蒂夫·沃兹尼亚克实际上提出了所谓的“咖啡测试”，而不是图灵测试，后者在很多方面都非常狭隘。咖啡测试很有趣：除非你开发出一个可以进入普通的、以前不为人知的美国家庭并以某种方式弄清楚如何煮一杯咖啡的系统，否则我们还没有解决通用人工智能。这听起来微不足道但同时又非常深刻，因为你正在解决大量不可知的一般问题，以便在一个陌生的家庭中煮一杯咖啡，你不知道东西会放在哪里，咖啡机是什么类型，他们有什么其他工具等。这是非常复杂的广义问题解决，涉及系统必须解决的众多类别的问题。因此，如果你想测试通用人工智能，我们可能需要这种形式的图灵测试，也许这就是我们需要去的地方。

JAMES MANYIKA: Steve Wozniak, the co-founder of Apple, has actually proposed what he calls the “coffee test” as opposed to Turing tests, which are very narrow in many respects. A coffee test is kind of fun: until you get a system that can enter an average and previously unknown American home and somehow figure out how to make a cup of coffee, we’ve not solved AGI. The reason why that sounds trivial but at the same time quite profound is because you’re solving a large number of unknowable and general problems in order to make that cup of coffee in an unknown home, where you don’t know where things are going to be, what type of coffee maker it is or other tools they have, etc. That’s very complex generalized problem-solving across numerous categories of problems that the system would have to do. Therefore, it may be that we need Turing tests of that form if you want to test for AGI, and maybe that’s where we need to go.

我应该指出另一个限制，即潜在问题不是算法问题，而是数据问题。这是一个大问题，往往会分裂人工智能社区。一种观点认为，这些机器的偏见可能比人类更少。你可以看很多例子，比如人类法官和保释决定，使用算法可以消除许多固有的人类偏见，包括人为的缺陷，甚至是一天中的时间偏见。招聘和晋升决策可能是另一个类似的领域，想想玛丽安·伯特兰 (Marianne Bertrand) 和森德希尔·穆莱纳坦 (Sendhil Mullainathan) 的研究，他们研究了不同种族群体提交相同简历求职时收到的回电差异。

I should point out the other limitation, which is the question of potential issues not so much in the algorithm, but in the data. This is a big question which tends to divide the AI community. One view is the idea that these machines are probably going to be less biased than humans. You can look at multiple examples, such as human judges and bail decisions where using an algorithm could take out many of the inherent human biases, including human fallibility and even time of day biases. Hiring and advancement decisions could be another similar area like this, thinking about Marianne Bertrand and Sendhil Mullainathan’s work looking at the difference in calls back received by different racial groups who submitted identical resumes for jobs.

马丁·福特：这是我在为本书进行的多次对话中谈到的问题。我们希望人工智能能够超越人类的偏见，但问题似乎总是在于，你用来训练人工智能系统的数据包含着人类的偏见，所以算法会拾取这些偏见。

MARTIN FORD: That’s something that has come up in a number of the conversations I’ve had for this book. The hope should be that AI can rise above human biases, but the catch always seems to be that that the data you’re using to train the AI system encapsulates human bias, so the algorithm picks it up.

詹姆斯·玛尼卡：确实如此，这是对偏见问题的另一种看法，即承认数据本身实际上可能存在很大偏差，包括数据收集、采样率（通过过度采样或欠采样）以及这对不同人群或不同类型的资料的系统意义。

JAMES MANYIKA: Exactly, that’s the other view of the bias question that recognizes that the data itself could actually be quite biased, both in its collection, the sampling rates—either through oversampling or undersampling—and what that means systematically, either to different groups of people or different kinds of profiles.

在贷款、警务和刑事司法案件中，普遍存在的偏见问题已经得到了相当惊人的体现，因此，在我们想要使用的任何数据集中，都可能存在大规模的偏见，其中许多偏见可能是无意的。Julia Angwin 和她在 ProPublica 的同事在他们的工作中强调了这种偏见，麦克阿瑟奖获得者 Sendhil Mullainathan 和他的同事也是如此。顺便说一句，这项研究最有趣的发现之一是，算法在数学上可能无法同时满足公平的不同定义，因此决定如何定义公平正成为一个非常重要的问题。

The general bias problem has been shown in quite a spectacular fashion in lending, in policing and criminal justice cases, and so in any dataset that we have want to use, we could have large-scale biases already built it, many likely unintended. Julia Angwin and her colleagues at ProPublica have highlighted such biases in their work, as has MacArthur Fellow Sendhil Mullainathan and his colleagues. One of the most interesting findings to come out of that work, by the way, is that algorithms may be mathematically unable to satisfy different definitions of fairness at the same time, so deciding how we will define fairness is becoming a very important issue.

我认为这两种观点都是正确的。一方面，机器系统可以帮助我们克服人类的偏见和易犯错误性，但另一方面，它们也可能带来更大的问题。这是我们需要努力克服的另一个重要限制。但在这里我们又开始取得进展了。我对 DeepMind 的 Silvia Chiappa 所做的开创性工作特别感到兴奋，她使用反事实公平和因果模型方法来解决公平和偏见问题。

I think both views are valid. On the one hand, machine systems can help us overcome human bias and fallibility, and yet on the other hand, they could also introduce potentially larger issues of their own. This is another important limitation we’re going to need to work our way through. But here again we are starting to make progress. I am particularly excited about the pioneering work that Silvia Chiappa at DeepMind is doing using counterfactual fairness and causal model approaches to tackle fairness and bias.

马丁·福特：那是因为数据直接反映了人们的偏见，对吧？如果数据是在人们正常行为时（使用在线服务或其他方式）收集的，那么数据最终会反映出他们的偏见。

MARTIN FORD: That’s because the data directly reflects the biases of people, right? If it’s collected from people as they’re behaving normally, using an online service or something, then the data is going to end up reflecting whatever biases they have.

詹姆斯·马尼卡：没错，但即使个人不一定有偏见，这实际上也可能是一个问题。我举个例子，你实际上不能责怪人类本身或他们自己的偏见，但这反而向我们展示了我们的社会是如何运作的，从而造成了这些挑战。以警务为例。例如，我们知道，一些社区的警力比其他社区更强，而根据定义，每当社区的警力更强时，就会收集到更多关于这些社区的数据以供算法使用。

JAMES MANYIKA: Right, but it can actually be a problem even if individuals aren’t necessarily biased. I’ll give you an example where you can’t actually fault the humans per se, or their own biases, but that instead shows us how our society works in ways that create these challenges. Take the case of policing. We know that, for example, some neighborhoods are more policed than others and by definition, whenever neighborhoods are more policed, there’s a lot more data collected about those neighborhoods for algorithms to use.

因此，如果我们选择两个社区，一个社区治安良好，另一个社区治安不佳（无论有意还是无意），那么这两个社区之间的数据采样差异将对犯罪预测产生影响。实际数据采集本身可能没有显示任何偏差，但由于一个社区采样过度，另一个社区采样不足，使用这些数据可能会导致有偏差的预测。

So, if we take two neighborhoods, one that is highly policed and one that is not—whether deliberately or not—the fact is that the data sampling differences across those two communities will have an impact on the predictions about crime. The actual collection itself may not have shown any bias, but because of oversampling in one neighborhood and undersampling in another, the use of that data could lead to biased predictions.

另一个欠采样和过采样的例子可以在贷款中看到。在这个例子中，它的作用是相反的，如果你有一个群体，因为他们使用信用卡和进行电子支付，所以有更多可用的交易，我们就会有更多的关于这个群体的数据。过采样实际上有助于这些群体，因为我们可以对他们做出更好的预测，而如果你有一个欠采样的群体，因为他们用现金支付，而且可用的数据很少，那么算法对这些群体的准确性可能会降低，因此在选择贷款时会更加保守，这实际上会影响最终的决策。我们在面部识别系统中也遇到了这个问题，Timnit Gebru、Joy Buolamwini 等人的工作已经证明了这一点。

Another example of undersampling and oversampling can be seen in lending. In this example, it works the other way, where if you have a population that has more available transactions because they’re using credit cards and making electronic payments, we have more data about that population. The oversampling there actually helps those populations, because we can make better predictions about them, whereas if you then have an undersampled population, because they’re paying in cash and there is little available data, the algorithm could be less accurate for those populations, and as a result, more conservative in choosing to lend, which essentially biases the ultimate decisions. We have this issue too in facial recognitions systems which has been demonstrated in the work of Timnit Gebru, Joy Buolamwini, and others.

这可能不是人类在开发算法时所存在的偏见，而是我们收集算法训练数据的方式引入了偏见。

It may not be the biases that any human being has in developing the algorithms, but the way in which we’ve collected the data that the algorithms are trained on that introduces bias.

马丁·福特：那么与人工智能相关的其他风险呢？最近备受关注的一个问题是超级智能可能带来的生存风险。您认为我们应该担心什么？

MARTIN FORD: What about other kinds of risks associated with AI? One issue that’s gotten a lot of attention lately is the possibility of existential risk from superintelligence. What do you think are the things we should legitimately worry about?

詹姆斯·马尼卡：有很多事情需要担心。我记得几年前，我们一群人，包括许多人工智能先驱和其他杰出人物，如伊隆·马斯克和斯图尔特·拉塞尔，在波多黎各开会讨论人工智能的进展以及需要更多关注的问题和领域。该小组最终在斯图尔特·拉塞尔发表的一篇论文中写到了一些问题，以及我们应该担心什么，并指出了在分析这些领域方面没有给予足够的关注和研究。自那次会议以来，需要担心的领域在过去几年里开始有所变化，但这些领域包括一切——包括安全问题。

JAMES MANYIKA: Well, there are lots of things to worry about. I remember a couple of years ago, a group of us, that included many of the AI pioneers and other luminaries, including the likes of Elon Musk and Stuart Russell, met in Puerto Rico to discuss progress in AI as well concerns and areas that needed more attention. The group ended up writing about what some of the issues are, in a paper that was published by Stuart Russell, and what we should worry about, and pointing out where there was not enough attention and research going into analyzing these areas. Since that meeting, the areas to worry about have begun to change a little bit in the last couple of years, but those areas included everything—including things like safety questions.

举个例子。如何阻止失控的算法？如何阻止失控的机器？我的意思不是终结者意义上的，而是狭义的算法，即做出错误解释、导致安全问题，甚至只是让人们心烦意乱的算法。为此，我们可能需要所谓的“大红按钮”，这是几个研究团队正在研究的东西。例如，DeepMind 在网格世界中的工作已经证明，许多算法理论上可以学习如何关闭自己的“关闭开关”。

Here is one example. How do you stop a runaway algorithm? How do you stop a runaway machine that gets out of control? I don’t mean in a Terminator sense, but even just in the narrow sense of an algorithm that is making wrong interpretations, leading to safety questions, or even simply upsetting people. For this we may need what has been referred to as the Big Red Button, something several research teams are working on DeepMind’s work with gridworlds, for example, has demonstrated that many algorithms could theoretically learn how to turn off their own “off-switches”.

另一个问题是可解释性。在这里，可解释性是一个用于讨论神经网络问题的术语：我们并不总是知道哪个特征或哪个数据集以某种方式影响了人工智能的决策或预测。这使得解释人工智能的决策变得非常困难，很难理解为什么它可能会做出错误的决定。当预测和决策具有可能影响生活的重大影响时，这可能非常重要，例如当人工智能用于刑事司法情况或贷款申请时，正如我们所讨论的那样。最近，我们看到了应对可解释性挑战的新技术的出现。一种有前途的技术是使用局部可解释模型不可知解释或 LIME。LIME 试图确定训练模型最依赖哪些特定数据集来做出预测。另一种有前途的技术是使用广义加性模型或 GAM。它们以加法方式使用单个特征模型，因此限制了特征之间的相互作用，因此可以随着特征的增加确定预测的变化。

Another issue is explainability. Here, explainability is a term used to discuss the problem that with neural networks: we don’t always know which feature or which dataset influenced the AI decision or prediction, one way or the other. This can make it very hard to explain an AI’s decision, to understand why it might be reaching a wrong decision. This can matter a great deal when predictions and decisions have consequential implications that may affect lives for example when AI is used in criminal justice situations or lending applications, as we’ve discussed. Recently, we’ve seen new techniques to get at the explainability challenge emerge. One promising technique is the use of Local-Interpretable-Model Agnostic Explanations, or LIME. LIME tries to identify which particular data sets a trained model relies on most to make a prediction. Another promising technique is the use of Generalized Additive Models, or GAMs. These use single feature models additively and therefore limit interactions between features, and so changes in predictions cane be determined as features are added.

我们应该更多地考虑的另一个领域是“检测问题”，即我们可能很难发现人工智能系统是否遭到恶意使用——可能是恐怖分子，也可能是犯罪分子。对于其他武器系统，比如核武器，我们有相当强大的检测系统。很难在无人知晓的情况下在世界上引爆核爆炸，因为你有地震测试、放射性监测和其他手段。对于人工智能系统，情况就没那么简单了，这引出了一个重要的问题：我们如何知道人工智能系统何时被部署？

Yet another area we should think about more is the “detection problem,” which is where we might find it very hard to even detect when there’s malicious use of an AI system—which could be anything from a terrorist to a criminal situation. With other weapons systems, like nuclear weapons, we have fairly robust detection systems. It’s hard to set off a nuclear explosion in the world without anybody knowing because you have seismic tests, radioactivity monitoring, and other things. With AI systems, not so much, which leads to an important question: How do we even know when an AI system is being deployed?

像这样的几个关键问题仍然需要大量的技术工作，我们必须取得进展，而不是每个人都逃避并专注于应用所带来的商业和经济效益。

There are several critical questions like this that still need a fair amount of technical work, where we must make progress, instead of everybody just running away and focusing on the upsides of applications for business and economic benefits.

这一切的希望在于，一些团体和实体正在涌现，并开始着手解决许多挑战。一个很好的例子就是人工智能伙伴关系。如果你看看伙伴关系的议程，你会发现很多问题正在被研究，关于偏见、关于安全，以及关于这类生存威胁的问题。另一个很好的例子是 Sam Altman、Jack Clarke 和 OpenAI 的其他人正在做的工作，旨在确保整个社会都能从人工智能中受益。

The silver lining of all this is that groups and entities are emerging and starting to work on many of these challenges. A great example is the Partnership on AI. If you look at the agenda for the Partnership, you’ll see a lot of these questions are being examined, about bias, about safety, and about these kinds of existential threat questions. Another great example is the work that Sam Altman, Jack Clarke and others at OpenAI are doing, which aims to make sure all of society benefits from AI.

目前，在这些问题上取得最大进展的实体和团体往往是能够吸引人工智能超级明星的地方，即使在 2018 年，这群超级明星也往往是一个相对较小的群体。希望随着时间的推移，这一群体会逐渐分散。我们还看到一些相对集中的人才流向拥有巨大计算能力和容量的地方，以及拥有大量数据独特访问权限的地方，因为我们知道这些技术受益于这些资源。问题是，在这个世界上，更多的进步往往流向拥有超级明星、数据可用和计算机能力可用的地方，你如何确保这些进步继续广泛地惠及所有人？

Right now, the entities and groups that are making the most progress on these questions have tended to be places that have been able to attract the AI superstars, which, even in 2018, tends to be a relatively small group. That will hopefully diffuse over time. We’ve also seen some relative concentrations of talent go to places that have massive computing power and capacity, as well as places that have unique access to lots of data, because we know these techniques benefit from those resources. The question is, in a world in which there’s a tendency for more progress to go to where the superstars are, and where the data is available, and where the computer capacity is available, how do you make sure this continues to be widely available to everybody?

马丁·福特：你对存在主义问题有什么看法？埃隆·马斯克和尼克·博斯特罗姆讨论了控制问题或协调问题。一种情况是，我们可以通过递归改进实现快速起飞，然后我们就会得到一台远离我们的超级智能机器。这是我们现在应该担心的事情吗？

MARTIN FORD: What do you think about the existential concerns? Elon Musk and Nick Bostrom talk about the control problem or the alignment problem. One scenario is where we could have a fast takeoff with recursive improvement, and then we’ve got a superintelligent machine that gets away from us. Is that something we should be worried about at this point?

詹姆斯·马尼卡：是的，有人应该担心这些问题——但不是每个人都应该担心，部分原因是我认为超级智能机器的出现时间还很遥远，而且出现的可能性相当低。但同样，从帕斯卡赌注的角度来看，有人应该考虑这些问题，但我不会让社会为存在主义问题而激动，至少现在不会。

JAMES MANYIKA: Yes, somebody should be worrying about those questions—but not everybody, partly because I think the time frame for a super intelligent machine is so far away, and because the probability of that is fairly low. But again, in a Pascal-wager like sense, somebody should be thinking about those questions, but I wouldn’t get society all whipped up about the existential questions, at least not yet.

我很喜欢像尼克博斯特罗姆这样聪明的哲学家对此进行思考，我只是认为这还不应该成为整个社会关注的重大问题。

I like the fact that a smart philosopher like Nick Bostrom is thinking about it, I just don’t think that it should be a huge concern for society as a whole just yet.

马丁·福特：我也是这么想的。如果一些智库愿意关注这些问题，这似乎是个好主意。但目前很难证明投入大量政府资源是合理的。而且无论如何，我们可能都不希望政客们深入研究这些事情。

MARTIN FORD: That’s also my thinking. If a few think tanks want to focus on these concerns, that seems like a great idea. But it would be hard to justify investing massive governmental resources at this point. And we probably wouldn’t want politicians delving into this stuff in any case.

詹姆斯·曼尼卡：不，这不应该是一个政治问题，但我也不同意那些认为这种情况发生的可能性为零以及没有人应该担心此事的人的说法。

JAMES MANYIKA: No, it shouldn’t be a political issue, but I also disagree with people who say that there is zero probability that this could happen and say that no-one should worry about it.

我们大多数人不应该担心这个问题。我认为我们应该更担心现在存在的更具体的问题，例如安全性、使用和误用、可解释性、偏见、经济和劳动力影响问题以及相关的转变。这些是更大、更现实的问题，将从现在开始影响社会，并在未来几十年内持续下去。

The vast majority of us shouldn’t be worried about it. I think that we should be more worried about these more specific questions that are here now, such as safety, use and misuse, explainability, bias, and the economic and workforce effects questions and related transitions. Those are the bigger, more real questions that are going to impact society beginning now and running over the next few decades.

马丁·福特：就这些问题而言，您认为监管是否有必要？政府是否应该介入并监管人工智能的某些方面，还是应该依靠行业自行解决问题？

MARTIN FORD: In terms of those concerns, do you think there’s a place for regulation? Should governments step in and regulate certain aspects of AI, or should we rely on industry to figure it out for themselves?

詹姆斯·马尼卡：我不知道监管应该采取什么形式，但有人应该考虑在这种新环境下如何监管。我认为我们目前还没有任何工具，也没有任何正确的监管框架。

JAMES MANYIKA: I don’t know what form regulation should take, but somebody should be thinking about regulation in this new environment. I don’t think that we’ve got any of the tools in place, any of the right regulatory frameworks in place at all right now.

所以，我的简单回答是，是的，应该有人考虑应该如何监管人工智能。但我认为监管不应该从这样的观点开始：其目标是阻止人工智能，重新盖上潘多拉魔盒，或者阻止这些技术的部署，并试图让时光倒流。

So, my simple answer would be yes, somebody should be thinking about what the regulation of AI should look like. But I think the regulation shouldn’t start with the view that its goal is to stop AI and put back the lid on a Pandora’s box, or hold back the deployment of these technologies and try and turn the clock back.

我认为这种说法是错误的，因为首先，妖怪已经从瓶子里出来了；但更重要的是，这些技术带来了巨大的社会和经济效益。我们可以更多地谈论我们的整体生产力挑战，而这些挑战正是这些人工智能系统可以提供帮助的。我们还有社会“登月”挑战，而人工智能系统可以帮助解决这些挑战。

I think that would be misguided because first of all, the genie is out of the bottle; but also, more importantly, there’s enormous societal and economic benefit from these technologies. We can talk more about our overall productivity challenge, which is something these AI systems can help with. We also have societal “moonshot” challenges that AI systems can help with.

因此，如果监管的目的是减缓或阻止人工智能的发展，那么我认为这是错误的，但如果监管的目的是考虑安全问题、隐私问题、透明度问题以及这些技术的广泛可用性问题，以便每个人都能从中受益——那么我认为这些都是人工智能监管应该考虑的正确的事情。

So, if regulation is intended to slow things down or stop the development of AI then I think that’s wrong, but if regulation is intended to think about questions of safety, questions of privacy, questions of transparency, questions around the wide availability of these techniques so that everybody can benefit from them—then I think those are the right things that AI regulation should be thinking about.

马丁·福特：我们来谈谈经济和商业方面的问题。我知道麦肯锡全球研究院已经发布了几份关于人工智能对工作和劳动力影响的重要报告。

MARTIN FORD: Let’s move on to the economic and business aspects of this. I know the McKinsey Global Institute has put out several important reports on the impact of AI on work and labor.

我已经写了很多关于这个话题的文章，我的上一本书中提出了一个观点，即我们正处于一场重大变革的前沿，这场变革可能会对劳动力市场产生巨大影响。你的看法是什么？我知道有不少经济学家认为这个问题被夸大了。

I’ve written quite a lot on this, and my last book makes the argument that we’re really on the leading edge of a major disruption that could have a huge impact on labor markets. What’s your view? I know there are quite a few economists who feel this issue is being overhyped.

詹姆斯·马尼卡：不，它并没有被过度炒作。我认为我们正处于转折点，即将进入一场新的工业革命。我认为这些技术将对企业产生巨大的、变革性的积极影响，因为它们的效率、对创新的影响、对预测和寻找问题新解决方案的影响，在某些情况下超越了人类的认知能力。根据我们在麦肯锡全球研究院的研究，我认为人工智能对企业的影响无疑是积极的。

JAMES MANYIKA: No, it is not overhyped. I think we’re on the cusp and we’re about to enter a new industrial revolution. I think these technologies are going to have an enormous, transformative and positive impact on businesses, because of their efficiency, their impact on innovation, their impact on being able to make predictions and to find new solutions to problems, and in some case go beyond human cognitive capabilities. The impact of AI on business to me, based on our research at MGI, is for the businesses undoubtedly positive.

对经济的影响也将是相当变革性的，主要是因为这将导致生产力提高，而生产力是经济增长的引擎。这一切都将发生在我们面临老龄化和其他影响的时期，这些影响将给经济增长带来阻力。人工智能和自动化系统以及其他技术将对生产力产生这种变革性和急需的影响，从长远来看，这将带来经济增长。这些系统还可以显著加速创新和研发，从而带来新产品和服务，甚至是改变经济的商业模式。

The impact on the economy is also going to be quite transformational too, mostly because this is going to lead to productivity gains, and productivity is the engine of economic growth. This will all take place at a time when we’re going to have aging and other effects that will create headwinds for economic growth. AI and automation systems, along with other technologies, are going to have this transformational and much-needed effect on productivity, which in the long term leads to economic growth. These systems can also significantly accelerate innovation and R&D, which leads to new products and services and even business models that will transform the economy.

我也对它对社会的影响持非常乐观的态度，因为它能够解决我之前提到的社会“登月”挑战。这可能是一个新项目或应用，它能对社会挑战产生新的见解，或提出根本性的解决方案，或导致突破性技术的发展。这可能是医疗保健、气候科学、人道主义危机或发现新材料。这是我和同事正在研究的另一个领域，很明显，从图像分类到自然语言处理和对象识别等人工智能技术可以在这些领域中做出巨大贡献。

I’m also quite positive about the impact on society in the sense of being able to solve the societal “moonshot” challenges I hinted at before. This could a new project or application that yields new insights into a societal challenge or proposes a radical solution or leads to the development of a breakthrough technology. This could be in healthcare, climate science, humanitarian crises or in discovering new materials. This is another area that my colleagues and I are researching where it’s clear that AI techniques from image classification to natural language processing and object identification can make a big contribution in many of these domains.

话虽如此，如果你说人工智能有利于商业、有利于经济增长，并有助于解决社会难题，那么最大的问题是——工作呢？我认为这是一个更加复杂的故事。但如果我要总结一下我对工作的看法，我会说，工作会减少，但也会增多。

Having said all of that, if you say AI is good for business, good for economic growth, and helps tackle societal moonshots, then the big question is—what about work? I think this is a much more mixed and complicated story. But I think if I were to summarize my thoughts about jobs, I would say there will be jobs lost, but also jobs gained.

马丁·福特：那么，您认为尽管会有大量工作岗位流失，但净影响将是积极的吗？

MARTIN FORD: So, you believe the net impact will be positive, even though a lot of jobs will be lost?

詹姆斯·马尼卡：虽然会有一些工作岗位流失，但也会有一些工作岗位增加。在“增加工作岗位”方面，工作岗位将来自经济增长本身以及由此产生的活力。工作需求永远存在，而且通过生产力和经济增长，存在一些机制，可以增加工作岗位并创造新的工作岗位。此外，短期至中期内，有多个工作需求驱动因素相对稳定，包括随着越来越多的人进入消费阶层，世界各地的繁荣程度不断提高等等。另一件将要发生的事情是所谓的“工作岗位发生变化”，这是因为这些技术将以许多有趣的方式补充工作，即使我们不能完全取代从事这项工作的人。

JAMES MANYIKA: While there will be jobs lost, there’ll also be jobs gained. In the “jobs gained” side of the story, jobs will come from the economic growth itself, and from the resulting dynamism. There’s always going to be demand for work, and there are mechanisms, through productivity and economic growth, that lead to the growth of jobs and the creation of new jobs. In addition, there are multiple drivers of demand for work that are relatively assured in the near- to mid-term, these include, again, rising prosperity around the world as more people enter the consuming class and so on. Another thing which will occur is something called “jobs changed,” and that’s because these technologies are going to complement work in lots of interesting ways, even when we don’t fully replace people doing that work.

在之前的自动化时代，我们已经看到了这三种观念的版本：失业、就业增加和就业发生变化。真正的争论是，所有这些因素的相对大小是什么，我们最终会走向何方？我们失业的人数会比就业增加的人数多吗？这是一个有趣的争论。

We’ve seen versions of these three ideas of jobs lost, jobs gained, and jobs changed before with previous eras of automation. The real debate is, what are the relative magnitudes of all those things, and where do we end up? Are we going to have more jobs lost than jobs gained? That’s an interesting debate.

麦肯锡全球研究所的研究表明，我们将取得领先，新增就业岗位将多于流失就业岗位；当然，这是基于围绕几个关键因素的一系列假设。由于无法做出预测，我们围绕所涉及的多个因素制定了各种情景，在我们的中点情景中，我们取得了领先。有趣的问题是，即使在一个就业岗位充足的世界里，需要解决的关键劳动力问题是什么，包括对工资等方面的影响以及所涉及的劳动力转型？就业和工资状况比对企业和经济的影响更为复杂，就增长而言，正如我所说，这显然是积极的。

Our research at MGI suggests that we will come out ahead, that there will be more jobs gained than jobs lost; this of course is based on a set of assumptions around a few key factors. Because it’s impossible to make predictions, we have developed scenarios around the multiple factors involved, and in our midpoint scenarios we come out ahead. The interesting question is, even in a world with enough jobs, what will be the key workforce issues to grapple with, including the effect on things like wages, and the workforce transitions involved? The jobs and wages picture is more complicated than the effect on business and the economy, in terms of growth, which as I said, is clearly positive.

马丁·福特：在我们讨论就业和工资之前，让我先谈谈你的第一点：对企业的积极影响。如果我是经济学家，我会立即指出，如果你看看最近的生产率数据，你会发现它们真的不是那么好——就宏观经济数据而言，我们还没有看到生产率有任何提高。事实上，与其他时期相比，生产率一直相当低迷。你是说在事情起飞之前会有一个滞后吗？

MARTIN FORD: Before we talk about jobs and wages, let me focus on your first point: the positive impact on business. If I were an economist, I would immediately point out that if you look at the productivity figures recently, they’re really not that great—we are not seeing any increases in productivity yet in terms of the macro-economic data. In fact, productivity has been pretty underwhelming, relative to other periods. Are you arguing that there’s just a lag before things will take off?

詹姆斯·马尼卡：麦肯锡全球研究所最近发​​布了一份关于这一问题的报告。生产率增长缓慢的原因有很多，其中一个原因是过去 10 年，我们经历了约 70 年来资本密集度最低的时期。

JAMES MANYIKA: We at MGI recently put out a report on this. There are a lot of reasons why productivity growth is sluggish, one reason being that in the last 10 years we’ve had the lowest capital intensity period in about 70 years.

我们知道，资本投资和资本密集度是推动生产力增长的因素之一。我们也知道需求的关键作用——包括麦肯锡全球研究院在内的大多数经济学家往往关注生产力的供给侧效应，而较少关注需求侧。我们知道，当需求大幅放缓时，即使生产效率再高，测算出来的生产力也不会很高。这是因为生产力测量有分子和分母：分子涉及增值产出的增长，这要求产出被需求所吸收。因此，无论技术进步如何，如果需求因某种原因而滞后，就会损害产出增长，从而降低生产力增长。

We know that capital investment, and capital intensity, are part of the things that you need to drive productivity growth. We also know the critical role of demand—most economists, including here at MGI, have often looked at the supply-side effects of productivity, and not as much at the demand side. We know that when you’ve got a huge slowdown in demand you can be as efficient as you want in production, and measured productivity still won’t be great. That’s because the productivity measurement has a numerator and a denominator: the numerator involves growth in value-added output, which requires that output is being soaked up by demand. So, if demand is lagging for whatever reason, that hurts growth in output, which brings down productivity growth, regardless of what technological advances there may have been.

马丁·福特：这是很重要的一点。如果技术进步加剧了不平等，压低了工资，那么它实际上就从普通消费者的口袋里掏走了钱，那么这可能会进一步抑制需求。

MARTIN FORD: That’s an important point. If advancing technology increases inequality and holds down wages, so it effectively takes money out of the pockets of average consumers, then that could dampen down demand further.

詹姆斯·马尼卡：哦，当然。需求点绝对至关重要，尤其是在发达经济体中，55% 到 70% 的需求是由消费者和家庭支出驱动的。你需要人们有足够的收入来消费所有生产的产品。需求是故事的重要组成部分，但我认为还有你提到的技术滞后问题。

JAMES MANYIKA: Oh, absolutely. The demand point is absolutely critical, especially when you’ve got advanced economies, where anywhere between 55% and 70% of the demand in those economies is driven by consumer and household spending. You need people earning enough to be able to consume the output of everything being produced. Demand is a big part of the story, but I think there is also the technology lag story that you mentioned.

对于您最初的问题，我有幸在 1999 年至 2003 年间与麦肯锡全球研究所的一位学术顾问、诺贝尔奖获得者鲍勃·索洛共事。我们当时正在研究 1990 年代末的最后一个生产力悖论。在 80 年代末，鲍勃提出了后来被称为“索洛悖论”的观察结果，即除了生产力数字之外，计算机随处可见。这个悖论最终在 90 年代末得到解决，当时我们有足够的需求来推动生产力增长，但更重要的是，当时我们的经济中有很大一部分部门（零售、批发等）最终采用了当时的技术：客户端-服务器架构、ERP 系统。这改变了他们的业务流程，推动了经济中很大一部分部门的生产力增长，最终产生了足够大的影响来推动国家生产力的发展。

To your original question, I had the pleasure between 1999 and 2003 to work with one of the academic advisors of the McKinsey Global Institute, Bob Solow, the Nobel laureate. We were looking at the last productivity paradox back in the late 1990s. In the late ‘80s, Bob had made the observation that became known as The Solow Paradox, that you could see computers everywhere except in the productivity numbers. That paradox was finally resolved in the late ‘90s, when we had enough demand to drive productivity growth, but more importantly, when we had very large sectors of the economy—retail, wholesale, and others—finally adopting the technologies of the day: client-server architectures, ERP systems. This transformed their business processes and drove productivity growth in very large sectors in the economy, which finally had a big enough effect to move the national productivity needle.

现在，如果你快进到我们今天所处的位置，我们可能会看到类似的东西，如果你看看当前的数字技术浪潮，无论是云计算、电子商务还是电子支付，我们随处可见它们，我们都把它们放在口袋里，但生产率增长已经好几年很低迷了。但如果你真的系统地衡量当今经济的数字化程度，看看当前的数字技术浪潮，令人惊讶的答案是：实际上，就资产、流程和人们如何使用技术而言，数字化程度并不高。我们甚至还没有在这些数字化评估中谈论人工智能或下一波技术。

Now if you fast-forward to where we are today, we may be seeing something similar in the sense that if you look at the current wave of digital technologies, whether we’re talking about cloud computing, e-commerce, or electronic payments, we can see them everywhere, we all carry them in our pockets, and yet productivity growth has been very sluggish for several years now. But if you actually systematically measure how digitized the economy is today, looking at the current wave of digital technologies, the surprising answer is: not so much, actually, in terms of assets, processes, and how people work with technology. And we are not even talking about AI yet or the next wave of technologies with these assessments of digitization.

你会发现，相对而言，数字化程度最高的行业是科技行业本身、媒体和金融服务业。从宏观角度来看，这些行业实际上相对较小，以占 GDP 或就业的份额来衡量，而相对而言，规模非常大的行业数字化程度并不高。

What you find is that the most digitized sectors—on a relative basis—are sectors like the tech sector itself, media and maybe financial services. And those sectors are actually relatively small in the grand scheme of things, measuring as a share of GDP or as a share of employment, whereas the very large sectors are, relatively speaking, not that digitized.

以零售业为例，请记住零售业是最大的行业之一。我们都对电子商务的前景和亚马逊的所作所为感到兴奋。但现在通过电子商务进行的零售量只有 10% 左右，而亚马逊占了这 10% 的很大一部分。但零售业是一个非常大的行业，拥有许多中小型企业。这已经告诉你，即使在零售业这个我们认为高度数字化的大型行业之一，实际上我们还没有取得太大的广泛进展。

Take a sector like retail and keep in mind that retail is one of the largest sectors. We all get excited by the prospect of e-commerce and what Amazon is doing. But the amount of retail that is now done through e-commerce is only about 10%, and Amazon is a large portion of that 10%. But retail is a very large sector with many, many small- and medium-sized businesses. That already tells you that even in retail, one of the large sectors which we’d think of as highly digitized, in reality, it turns out we really haven’t yet made much widespread progress yet.

因此，我们可能正在经历另一轮索洛悖论。除非我们让这些非常大的行业高度数字化，并在业务流程中使用这些技术，否则我们不会看到足够的成果来推动全国生产力的发展。

So, we may be going through another round of the Solow paradox. Until we get these very large sectors highly digitized and using these technologies across business processes, we won’t see enough to move the national needle on productivity.

马丁·福特：那么，您的意思是，我们在全球范围内甚至还没有开始看到人工智能和先进自动化形式的影响？

MARTIN FORD: So, you’re saying that globally we haven’t even started to see to the impact of AI and advanced forms of automation yet?

詹姆斯·马尼卡：还没有。这又引出了另一个值得一提的观点：我们实际上需要的生产力增长甚至比我们想象的还要多，而人工智能、自动化和所有这些数字技术对于推动生产力增长和经济增长都至关重要。

JAMES MANYIKA: Not yet. And that gets to another point worth making: we’re actually going to need productivity growth even more than we can imagine, and AI, automation and all these digital technologies are going to be critical to driving productivity growth and economic growth.

为了解释原因，让我们看看过去 50 年的经济增长，看看 G20 国家（占全球 GDP 的 90% 多一点），我们有数据的过去 50 年，也就是 1964 年至 2014 年，平均经济 GDP 增长率为 3.5%。这是这些国家的平均 GDP 增长率。如果你做经典的增长分解和增长核算工作，就会发现 GDP 和经济增长来自两个因素：一个是生产力增长，另一个是劳动力供应的扩张。

To explain why, let’s look at the last 50 years of economic growth, and you look at that for the G20 countries (which make up a little more than 90% of global GDP), the average economic GDP growth over the last 50 years where we have the data, so between 1964 and 2014, was 3.5%. This was the average GDP growth across those countries. If you do classic growth decomposition and growth accounting work, it shows that GDP and economic growth comes from two things: one is productivity growth, and the other is expansions in the labor supply.

在过去 50 年中，我们平均 GDP 增长了 3.5%，其中 1.7% 来自劳动力供应的扩张，另外 1.8% 来自这 50 年的生产率增长。展望未来 50 年，由于老龄化和其他人口因素的影响，劳动力供应扩张带来的增长将从过去 50 年的 1.7% 急剧下降到 0.3% 左右。

Of the 3.5% of average GDP growth we’ve had in the last 50 years, 1.7% has come from expansions in the labor supply, and the other 1.8% has come from productivity growth over those 50 years. If you look to the next 50 years, the growth from expansions in the labor supply is going to come crashing down from the 1.7% that it’s been the last 50 years to about 0.3%, because of aging and other demographic effects.

因此，这意味着在未来 50 年，我们将比过去 50 年更加依赖生产力增长。除非我们的生产力大幅提高，否则经济增长将出现下滑。如果我们认为生产力增长目前对我们目前的增长至关重要，事实也确实如此，那么如果我们仍然希望实现经济增长和繁荣，那么在未来 50 年，生产力增长将更加重要。

So that means that in the next 50 years we’re going to rely even more than we have in the past 50 years on productivity growth. And unless we get big gains in productivity, we’re going to have a downdraft in economic growth. If we think productivity growth matters right now for our current growth, which it does, it’s going to matter even more for the next 50 years if we still want economic growth and prosperity.

马丁·福特：这有点触及经济学家罗伯特·戈登的观点，即未来经济可能不会有太大的增长。（Robert Gordon’s 2017 book The Rise and Fall of American Growth, offers a very pessimistic view of future economic growth in the United States）

MARTIN FORD: This is kind of touching on the economist Robert Gordon’s argument that may be there’s not going to be much economic growth in the future. (Robert Gordon’s 2017 book The Rise and Fall of American Growth, offers a very pessimistic view of future economic growth in the United States)

詹姆斯·马尼卡：鲍勃·戈登表示，经济可能不会增长，但他也质疑我们是否会有足够大的创新，与电气化等类似的东西相比，以真正推动经济增长。他怀疑是否会有像电力和过去的其他一些技术一样大的创新。

JAMES MANYIKA: While Bob Gordon’s saying there may not be economic growth, he’s also questioning whether we’re going to have big enough innovations, comparable to electrification and other things like that, to really drive economic growth. He’s skeptical that there’s going to be anything as big as electricity and some of the other technologies of the past.

马丁·福特：但是希望人工智能成为下一个热门话题？

MARTIN FORD: But hopefully AI is going to be that next thing?

詹姆斯·马尼卡：我们希望如此！它当然是像电力一样的通用技术，从这个意义上讲，应该会造福于多种活动和经济部门。

JAMES MANYIKA: We hope it will be! It is certainly a general-purpose technology like electricity, and in that sense should benefit multiple activities and sectors of the economy.

马丁·福特：我想进一步谈谈麦肯锡全球研究所关于工作和工资变化的报告。您能否详细介绍一下您编写的各种报告以及您的总体发现？您使用什么方法来确定某个特定工作是否可能被自动化，以及有多少比例的工作面临风险？

MARTIN FORD: I want to talk more about The McKinsey Global Institute’s reports on what’s happening to work and wages. Could you go into a bit more detail about the various reports you’ve generated and your overall findings? What methodology do you use to figure out if a particular job is likely to be automated and what percentage of jobs are at risk?

詹姆斯·曼尼卡：我们将其分为三个部分：“失业”、“就业发生变化”以及“就业增加”，因为每个途径都有值得讨论的内容。

JAMES MANYIKA: Let’s take this in three parts: “jobs lost,” “jobs changed,” and then “jobs gained,” because there’s something to be said about each of these pathways.

关于“失业”的问题，已经有很多研究和报告，这已经成为一种猜测就业问题的小产业。在麦肯锡全球研究院，我们认为我们采取的方法在两个方面略有不同。一是我们进行了基于任务的分解，所以我们从任务开始，而不是从整个职业开始。我们使用各种来源，包括 O*NET 数据集和我们通过查看任务获得的其他数据集，研究了超过 2,000 项任务和活动。然后，美国劳工统计局跟踪了大约 800 个职业；因此，我们将这些任务映射到实际职业中。

In terms of “jobs lost,” there’s been lots of research and reports, and it’s become a cottage industry speculating on the jobs question. At MGI the approach we’ve taken we think is a little bit different in two ways. One is that we’ve conducted a task-based decomposition, and so we’ve started with tasks, as opposed to starting with whole occupations. We’ve looked at something like over 2,000 tasks and activities using a variety of sources, including the O*NET dataset, and other datasets that we’ve got by looking at tasks. Then, the Bureau of Labor Statistics in the US tracks about 800 occupations; so, we mapped those tasks into the actual occupations.

我们还研究了执行这些任务所需的 18 种不同能力，我所说的能力包括认知能力、感知能力和完成这些任务所需的身体运动技能等所有能力。然后，我们试图了解现在的技术在多大程度上可以自动化并执行这些相同的能力，然后我们可以将这些能力映射到我们的任务上，并展示机器可以执行哪些任务。我们研究了所谓的“目前已证明的技术”，我们在这里区分的是已经在实验室或实际产品中实际证明的技术，而不仅仅是假设的技术。通过研究这些“目前已证明的技术”，我们可以在典型的采用和传播率的基础上，展望未来十五年左右的发展。

We’ve also looked at 18 different kinds of capabilities required to perform these tasks, and by capabilities, I’m talking everything from cognitive capabilities to sensory capabilities, to physical motor skills that are required to fulfill these tasks. We’ve then tried to understand to what extent technologies are now available to automate and perform those same capabilities, which then we can map back to our tasks and show what tasks machines can perform. We’ve looked at what we’ve called “currently demonstrated technology,” and what we’re distinguishing there is technology that has actually been demonstrated, either in a lab or in an actual product, not just something that’s hypothetical. By looking at these “currently demonstrated technologies,” we can provide a view into the next decade and a half or so, given typical adoption and diffusion rates.

通过观察所有这些，我们得出结论，在美国经济的任务层面，大约 50% 的活动（不是工作，而是任务，强调这一点很重要）原则上是可以实现自动化的。

By looking at all this, we have concluded that on a task level in the US economy, roughly about 50% of activities—not jobs, but tasks, and it’s important to emphasize this—that people do now are, in principle, automatable.

马丁·福特：您是说，基于我们现有的技术，现在一半工人的工作都可以实现自动化？

MARTIN FORD: You’re saying that half of what workers do could conceivably be automated right now, based on technology we already have?

詹姆斯·马尼卡：目前，基于目前已证实的技术，实现 50% 活动的自动化在技术上是可行的。但还有一些单独的问题，例如，这些可自动化的活动如何映射到整个职业中？

JAMES MANYIKA: Right now, it is technically feasible to automate 50% of activities based on currently demonstrated technologies. But there are also separate questions, like how do those automatable activities then map into whole occupations?

因此，当我们再将目光转向职业时，我们实际上发现只有大约 10% 的职业有超过 90% 的组成任务可以实现自动化。请记住，这是一个任务数字，而不是工作数字。我们还发现大约 60% 的职业有大约三分之一的组成活动可以实现自动化——当然，这种组合因职业而异。这个 60-30 已经告诉你，技术将补充或增强的职业比将被取代的职业要多得多。这导致了我之前提到的“工作岗位发生变化”现象。

So, when we then map back into occupations, we actually find that only about 10% of occupations have more than 90% of their constituent tasks automatable. Remember this is a task number, not a jobs number. We also find that something like 60% of occupations have about a third of their constituent activities automatable—this mix of course varies by occupation. This 60-30 already tells you that many more occupations will be complemented or augmented by technologies than will be replaced. This leads to the “jobs changed” phenomena I mentioned earlier.

马丁·福特：我记得您的报告发表时，媒体对其进行了非常正面的报道——认为由于大多数工作岗位中只有一部分会受到影响，所以我们不必担心失业。但是，如果您有三名员工，而他们每人三分之一的工作都实现了自动化，那么这是否会导致整合，即三名员工变成两名员工？

MARTIN FORD: I recall that when your report was published, the press put a very positive spin on it—suggesting that since only a portion of most jobs will be impacted, we don’t need to worry about job losses. But if you had three workers, and a third of each of their work was automated, couldn’t that lead to consolidation, where those three workers become two workers?

詹姆斯·马尼卡：当然，这就是我接下来要讲的内容。这是一个任务组合论证。它最初可能会给你一些适度的数字，但随后你开始意识到工作可以以许多有趣的方式重新配置。

JAMES MANYIKA: Absolutely, that’s where I was going to go next. This is a task composition argument. It might give you modest numbers initially, but then you start to realize that work could be reconfigured in lots of interesting ways.

例如，你可以合并和整合。也许临界点并不是你需要将一个职业中的所有任务都自动化；相反，也许当你接近 70% 的任务可以自动化时，你可能会说，“让我们将工作和工作流程整合在一起。”因此，最初的计算可能从适中的数字开始，但当你重新组织和整合工作时，受影响的工作数量开始变大。

For instance, you can combine and consolidate. Maybe the tipping point is not that you need all of the tasks in an occupation to be automatable; rather, maybe when you get close to say, 70% of the tasks being automatable, you may then say, “Let’s just consolidate and reorganize the work and workflow altogether.” So, the initial math may begin with modest numbers, but when you reorganize and consolidate the work, the number of impacted jobs start to get bigger.

然而，我们在麦肯锡全球研究院的研究中还考虑了另一组考虑因素，我们认为其他一些关于自动化问题的评估中没有考虑到这些因素。到目前为止，我们所描述的只是询问技术可行性问题，这给了你 50% 的数字，但这实际上只是你需要问的五个问题中的第一个。

However, there is yet another set of considerations that we’ve looked at in our research at MGI which we think have been missing in some of the other assessments on the automation question. Everything that we have described so far is simply asking the technical feasibility question, which gives you those 50% numbers, but that is really only the first of about five questions you need to ask.

第二个问题与开发和部署这些技术的成本有关。显然，仅仅因为某件事在技术上可行，并不意味着它就会实现。

The second question is around the cost of developing and deploying those technologies. Obviously, just because something’s technically feasible, doesn’t mean it will happen.

以电动汽车为例。事实证明我们可以制造电动汽车，事实上，这在 50 多年前就是可行的，但它们究竟是什么时候出现的？当购买、维护、充电等成本变得足够合理，消费者愿意购买它们，公司愿意部署它们时。这只是最近才发生的事情。

Look at electric cars. It’s been demonstrated we could build electric cars, and in fact that was a feasible thing to do more than 50 years ago, but when did they actually show up? When the costs of buying it, maintaining it, charging it, etc., became reasonable enough that consumers wanted to buy them and companies want to deploy them. That’s only happened very recently.

因此，部署成本显然是一个重要的考虑因素，而且会有很大差异，这取决于你谈论的是替代体力劳动的系统还是替代认知劳动的系统。通常，当你替代认知劳动时，它主要是软件和标准计算平台，因此边际成本经济学可以很快下降，所以成本不会很高。

So, the cost of deployment is clearly an important consideration and will vary a lot, depending on whether you’re talking about systems that are replacing physical work, versus systems that are replacing cognitive work. Typically, when you’re replacing cognitive work, it’s mostly software and a standard computing platform, so the marginal cost economics can come down pretty fast, so that doesn’t cost very much.

另一方面，如果你要取代体力劳动，那么你需要建造一台有移动部件的物理机器；而这些东西的经济成本虽然会下降，但不会像软件那样下降得那么快。因此，部署成本是第二个重要的考虑因素，这开始减慢部署速度，而部署速度最初可能只是通过简单考虑技术可行性而提出的。

If you’re replacing physical work, on the other hand, then you need to build a physical machine with moving parts; and the economics of those things, while they’ll come down, they’re not going to come down as fast as where things are just software. So, the cost of deployment is the second important consideration, which then starts to slow down deployment rates that might initially be suggested by simply looking at technical feasibility.

第三个考虑因素是劳动力市场需求动态，考虑劳动力质量和数量，以及与之相关的工资。让我通过两种不同的工作来说明这一点。我们来看看会计和园丁。首先让我们看看这些考虑因素在这些职业中如何发挥作用。

The third consideration is labor-market demand dynamics, taking into account labor quality and quantity, as well as the wages associated with that. Let me illustrate this by thinking in terms of two different kinds of jobs. We’ll look at an accountant, and we’ll look at a gardener. First let’s see how these considerations could play out in these occupations.

首先，从技术上讲，会计的大部分工作（主要是数据分析、数据收集等）实现自动化比较容易，而园丁的工作（主要是在高度非结构化的环境中进行的体力劳动）实现自动化则比较困难。在这种环境中，事物并不完全按照你希望的方式排列（例如在工厂中），而且可能会出现无法预见的障碍。因此，我们的第一个问题是，实现这些任务自动化的技术难度已经远远高于会计。

First, it is technically easier to automate large portions of what the accountant does, mostly data analysis, data gathering, and so forth, whereas it’s still technically harder to automate what a gardener does, which is mostly physical work in a highly unstructured environment. Things in these kinds of environments aren’t quite lined up exactly where you want them to be—as they would be in a factory, for example, and there’s unforeseen obstacles that can be in the way. So, the degree of technical difficulty of automating those tasks, our first question, is already far higher than your accountant.

然后我们来考虑第二个问题：部署系统的成本，这又回到我刚才提出的论点。对于会计来说，这需要在标准计算平台上运行边际成本接近零的软件。对于园丁来说，它是一台有许多活动部件的物理机器。部署物理机器的成本经济学总是比自动化会计的软件更昂贵——即使成本下降，而机器人的成本正在下降。

Then we get to the second consideration: the cost of deploying the system, which goes back to the argument I just made. In the case of the accountant, this requires software with near zero-marginal cost economics running on a standard computing platform. With the gardener, it’s a physical machine with many moving parts. The cost economics of deploying a physical machine is always going to be—even as costs come down, and they are coming down for robotic machines—more expensive than the software to automate an accountant.

现在我们来谈谈第三个关键考虑因素，即劳动力的数量和质量以及工资动态。在这里，它再次倾向于会计的自动化，而不是园丁的自动化。为什么？因为在美国，我们平均每小时付给园丁 8 美元，而我们每小时付给会计的工资是 30 美元。会计自动化的激励已经远远高于园丁自动化的激励。在我们解决这个问题的过程中，我们开始意识到，从技术和经济角度来看，其中一些低薪工作实际上可能更难实现自动化。

Now to our third key consideration, that is the quantity and quality of labor, and the wage dynamics. Here again it favors an automating the accountant, rather than automating the gardener. Why? Because we pay a gardener, on average in the United States, something like $8 an hour; whereas we pay an accountant something like $30 an hour. The incentive to automate the accountant is already far higher than the incentive to automate the gardener. As we work our way through this, we start to realize that it may very well be that some of these low-wage jobs may actually be harder to automate, from both a technical and economic perspective.

马丁·福特：这对大学毕业生来说真是个坏消息。

MARTIN FORD: This sounds like really bad news for university graduates.

詹姆斯·马尼卡：别急。通常的区分是高工资与低工资；或高技能与低技能。但我真的不知道这种区分是否有用。

JAMES MANYIKA: Not so fast. Often the distinction that’s made is high wage versus low wage; or high skill versus low skill. But I really don’t know if that’s a useful distinction.

我想指出的是，可能实现自动化的活动与传统的工资结构或技能要求概念并不完全一致。如果正在进行的工作主要是数据收集、数据分析或在高度结构化的环境中进行体力劳动，那么大部分工作都可能实现自动化，无论传统上是高薪还是低薪、高技能还是低技能。另一方面，很难实现自动化的活动也跨越了工资结构和技能要求，包括需要判断或管理人员的任务，或在高度非结构化和意外环境中进行的体力劳动。因此，许多传统的低薪和高薪工作都面临自动化，这取决于活动，但许多其他传统的低薪和高薪工作也可能受到保护，不受自动化的影响。

The point I want to make is that the activities likely to be automated don’t line up neatly with traditional conceptions of wages structures or skills requirements. If the work that’s being done looks like mostly data collection, data analysis, or physical work in a highly structured environment, then much of that work is likely to be automated, whether it’s traditionally been high wage or low wage, high skill or low skill. On the other hand, activities that are very difficult to automate also cut across wage structures and skills requirements, including tasks that require judgment or managing people, or physical work in highly unstructured and unexpected environments. So many traditionally low wage and high wage jobs are exposed to automation, depending on the activities, but also many other traditionally low wage and high wages jobs may be protected from automation.

我想确保我们也能涵盖所有不同的因素。第四个关键考虑因素与包括劳动力替代在内的益处有关。有些领域会实现自动化，但这并不是因为你想节省劳动力成本，而是因为你实际上得到了更好的结果，甚至是超人的结果。在这些领域，你可以获得更好的感知或预测，这是人类能力无法实现的。最终，自动驾驶汽车可能会成为一个例子，一旦它们达到比人类驾驶更安全、犯错更少的程度。当你开始超越人类能力并看到性能改进时，这确实可以加快部署和采用的商业案例。

I want to make sure we cover all the different factors at play here, as well. The fourth key consideration has to do with benefits including and beyond labor substitution. There are going to be some areas where you’re automating, but it’s not because you’re trying to save money on labor, it is because you’re actually getting a better result or even a superhuman outcome. Those are places where you’re getting better perception or predictions that you couldn’t get with human capabilities. Eventually, autonomous vehicles will likely be an example of this, once they reach the point where they are safer and commit fewer errors than humans driving. When you start to go beyond human capabilities and see performance improvements, that can really speed up the business case for deployment and adoption.

第五个考虑因素可以称为社会规范，这是一个广义的术语，涵盖了我们可能遇到的潜在监管因素和社会接受因素。无人驾驶汽车就是一个很好的例子。今天，我们已经完全接受了这样一个事实，即大多数商用飞机只有不到 7% 的时间由真正的飞行员驾驶。其余时间，飞机都在自行飞行。没有人真正关心飞行员的情况，即使这个比例下降到 1%，是因为没有人能看到驾驶舱内部。门关着，我们坐在飞机上。我们知道里面有飞行员，但是否知道他们在飞行并不重要，因为我们看不到。而对于无人驾驶汽车，人们常常感到害怕的是，你可以真正地看着驾驶座，但那里没有人；汽车在自行行驶。

The fifth consideration could be called societal norms, which is a broad term for the potential regulatory factors and societal acceptance factors we may encounter. A great example of this can be seen in driverless vehicles. Today, we already fully accept the fact that most commercial planes are only being piloted by an actual pilot less than 7% of the time. The rest of the time, the plane is flying itself. The reason no-one really cares about the pilot situation, even if it goes down to 1%, is because no-one can see inside the cockpit. The door is closed, and we’re sitting on a plane. We know there’s a pilot in there, but whether we know that they’re flying or not doesn’t matter because we can’t see. Whereas with a driverless car, what often freaks people out is the fact that you can actually look in the driver’s seat and there’s no-one there; the car’s moving on its own.

目前，有很多研究正在研究人们与机器互动的社会接受度或舒适度。麻省理工学院等机构正在研究不同年龄组、不同社会环境和不同国家/地区的社会接受度。例如，在日本等地，在社会环境中使用物理机器比在其他一些国家/地区更容易接受。我们还知道，例如，不同年龄组对机器的接受度或多或少不同，并且可能因不同的环境或设置而异。如果我们转到医疗环境，医生进入后面的房间使用机器，在视线之外，然后回来告诉你诊断结果——这样可以吗？我们大多数人都会接受这种情况，因为我们实际上并不知道医生在后面的房间里发生了什么。但是，如果屏幕进入你的房间，诊断结果突然出现，而没有人在场与你交谈，我们会感到舒服吗？我们大多数人可能不会。所以，我们知道社会环境会影响社会接受度，这也会影响未来这些技术的采用和应用。

There’s a lot of research going on now looking at people’s social acceptance or comfort with interacting with machines. Places like MIT are looking at social acceptance across different age groups, across different social settings, and across different countries. For example, in places like Japan, having a physical machine in a social environment is a bit more acceptable than in some other countries. We also know that, for example, different age groups are more or less accepting of machines, and it can vary depending on different environments or settings. If we move to a medical setting, with a doctor who goes into the back room to use a machine, out of view, and then just comes back with your diagnosis—is that okay? Most of us would accept that situation, because we don’t actually know what happened in the back room with the doctor. But if a screen wheels into your room and a diagnosis just pops up without a human there to talk you through it, would we be comfortable with that? Most of us probably wouldn’t be. So, we know that social settings affect social acceptance, and that this is going to also affect where see these technologies adopted and applied in the future.

马丁·福特：但归根结底，这对整体就业意味着什么？

MARTIN FORD: But at the end of the day, what does this mean for jobs across the board?

詹姆斯·曼尼卡：好吧，关键在于，当你考虑完这五个关键因素后，你就会开始意识到，自动化的速度和程度，以及将会减少的工作岗位的范围，实际上是一个更加微妙的图景，可能会因职业和地点的不同而有所差异。

JAMES MANYIKA: Well, the point is that as you work your way through these five key considerations, you start to realize that the pace and extent of automation, and indeed the scope of the jobs that are going to decline, is actually a more deeply nuanced picture that’s likely to vary from occupation to occupation and place to place.

在麦肯锡全球研究所的上一份报告中，我们考虑了我刚才描述的因素，特别是工资、成本和可行性，并制定了多种情景。我们的中点情景表明，到 2030 年，全球可能会失去多达 4 亿个工作岗位。这是一个惊人的数字，但占全球劳动力的比例约为 15%。不过，考虑到我们一直在讨论的劳动力市场动态，尤其是工资，发达国家的失业率将高于发展中国家。

In our last report at MGI, which considered the factors I just described, and in particular considered wages, costs and feasibility, we developed a number of scenarios. Our midpoint scenario suggests that as many as 400 million jobs could be lost globally by 2030. This is an alarmingly large number, but as a share of the global labor force that is about 15%. It will be higher, though, in advanced countries than in developing countries, given labor-market dynamics, especially wages, that we’ve been discussing.

然而，所有这些情景显然都取决于技术是否加速发展，而这是有可能的。如果确实如此，那么我们对“目前已展示的技术”的假设就不成立了。此外，如果部署成本下降的速度比我们预期的还要快，情况也会发生变化。这就是为什么我们在实际构建的情景中对会有多少工作岗位流失给出了如此广泛的范围。

However, all these scenarios are obviously contingent on whether the technology accelerates even faster, which it could. If it did, then our assumption about “currently demonstrated technology” would be out of the window. Further, if the costs of deploying come down even faster than we anticipate, that would also change things. That’s why we’ve got these wide ranges in the scenarios that we’ve actually built for how many jobs would be lost.

马丁·福特：那么“增加就业机会”方面呢？

MARTIN FORD: What about the “jobs gained” aspect?

詹姆斯·马尼卡：就业岗位增加这一方面很有意思，因为我们知道，只要经济不断增长、充满活力，就业岗位和工作需求就会增加。过去 200 年的经济增长史就是这样的，经济充满活力、不断发展，私营部门也充满活力。

JAMES MANYIKA: The “jobs gained” side of things is interesting because we know that whenever there’s a growing and dynamic economy, there will be growth in jobs and demand for work. This has been the history of economic growth for the last 200 years, where you’ve got vibrant, growing economies with a dynamic private sector.

如果我们展望未来 20 年左右，会发现一些相对确定的就业需求驱动因素。其中之一是全球繁荣的不断增长，因为越来越多的人进入消费阶层，对产品和服务有需求。另一个因素是老龄化；我们知道老龄化将产生对某些工作类型的大量需求，从而导致大量工作和职业的增长。现在还有一个单独的问题，即这些工作是否会变成高薪工作，但我们知道对护理工作和其他工作的需求将会增加。

If we look ahead to the next 20 years or so, there are some relatively assured drivers of demand for work. One of them is rising global prosperity as more people around the work enter the consuming class and demand products and services. Another is aging; and we know that aging is going to create a lot of demand for certain kinds of work that will lead to growth in a whole host of jobs and occupations. Now there’s a separate question as to whether those will turn into well-paying jobs or not, but we know that the demand for care work and other things is going to go up.

在麦肯锡全球研究所，我们还研究了其他催化剂，比如我们是否会加大对气候变化的适应力度，改造我们的系统和基础设施——这可能会推动对工作的需求超出目前的进程和速度。我们还知道，如果美国等社会最终齐心协力，关注基础设施的增长，并对基础设施进行投资，那么这也将推动对工作的需求。因此，工作机会的一个来源是不断增长的经济和这些对工作需求的具体驱动因素。

At MGI we’ve also looked at other catalysts, like whether we’re going to ramp up adaptation for climate change, retrofitting our systems and our infrastructure—which could drive demand for work above and beyond current course and speed. We also know that if societies like the United States and others finally get their act together to look at infrastructure growth, and make investments in infrastructure, then that’s also going to drive demand for work. So, one place where work’s going to come from is a growing economy and these specific drivers of demand for work.

另一大类工作将来自于我们实际上将发明以前不存在的新职业这一事实。我们在麦肯锡全球研究院进行的一项有趣分析——这是由我们的一位学术顾问、哈佛大学的迪克·库珀 (Dick Cooper) 提出的——是研究劳工统计局的数据。该局通常跟踪大约 800 个职业，底部总会有一行称为“其他”。这组称为“其他”的职业通常反映在当前测量期间尚未定义且不存在的职业，因此劳工统计局没有为它们设立类别。现在，如果您查看 1995 年的劳工统计局名单，网页设计师将属于“其他”类别，因为之前没有想象过，所以没有对其进行分类。有趣的是，“其他”类别是增长最快的职业类别，因为我们不断发明以前不存在的职业。

Another whole set of jobs are going to come from the fact that we’re actually going to invent new occupations that didn’t previously exist. One of the fun analyses we did at MGI—and this was prompted by one of our academic advisors, Dick Cooper at Harvard—was to look at the Bureau of Labor Statistics. This typically tracks about 800 occupations, and there’s always a line at the bottom called “Other.” This bucket of occupations called “Other” typically reflects occupations that in the current measurement period have not yet been defined and didn’t exist, so the Bureau doesn’t have a category for them. Now, if you had looked at the Labor Statistics list in 1995, a web designer would have been in the “Other” category because it hadn’t been imagined previously, and so it hadn’t been classified. What’s interesting is that the “Other” category is the fastest-growing occupational category because we’re constantly inventing occupations that didn’t exist before.

马丁·福特：我经常听到这种说法。例如，就在 10 年前，涉及社交媒体的工作还不存在。

MARTIN FORD: This is an argument that I hear pretty often. For example, just 10-years ago, jobs that involve social media did not exist.

詹姆斯·马尼卡：没错！如果你看看美国的十年历史，至少有 8% 到 9% 的工作岗位是前一时期不存在的——因为我们创造了它们并发明了它们。这将成为另一个就业来源，我们甚至无法想象它们会是什么，但我们知道它们会存在。一些人推测这一类别将包括新类型的设计师，以及排除机器和机器人故障和管理的人。这组未定义的新工作将成为工作的另一个驱动力。

JAMES MANYIKA: Exactly! If you look at 10-year periods in the United States, at least 8% to 9% of jobs are jobs that didn’t exist in the prior period—because we’ve created them and invented them. That’s going to be another source of jobs, and we can’t even imagine what those will be, but we know they’ll be there. Some people have speculated that category will include new types of designers, and people who trouble shoot and manage machines and robots. This undefined, new set of jobs will be another driver of work.

当我们研究新增就业岗位的类型，并考虑这些不同的动态因素时，除非经济下滑并出现大规模停滞，否则新增就业岗位的数量足以弥补失业人数。当然，除非我们身处的某些变量发生灾难性变化，例如这些技术的开发和采用显著加速，否则我们最终将面临大规模经济停滞。如果这些因素有任何组合，那么是的，我们最终将失去比新增更多的就业岗位。

When we’ve looked at the kind of the jobs gained, and considered these different dynamics, then unless the economy tanks and there’s massive stagnation, the numbers of jobs gained are large enough to more than make up for the jobs lost. Unless, of course, some of variables change catastrophically underneath us, such as a significant acceleration in the development and adoption of these technologies, or we end up with massive economic stagnation. Any combination of those things and then yes, we’ll end up with more jobs lost than jobs gained.

马丁·福特：好的，但如果你看一下就业统计数据，难道不是大多数工人都从事非常传统的工作吗，比如收银员、卡车司机、护士、教师、医生或办公室职员？这些都是 100 年前就存在的工作类别，而且绝大多数劳动力仍然从事这些工作。

MARTIN FORD: Ok, but if you look at the employment statistics, aren’t most workers employed in pretty traditional areas, such as cashiers, truck drivers, nurses, teachers, doctors or office workers? These are all job categories that were here 100 years ago, and that’s still where the vast majority of the workforce is employed.

詹姆斯·马尼卡：是的，经济仍然由这些职业的很大一部分组成。虽然其中一些职业会减少，但很少有职业会完全消失，而且消失的速度肯定不会像某些人预测的那样快。实际上，我们研究的一件事是，在过去 200 年里，我们看到了最大规模的就业岗位减少。例如，我们研究了美国制造业发生了什么，以及从农业到工业化的转变。我们研究了不同国家 20 个不同的大规模就业岗位减少情况，以及每个国家的情况，并与自动化和人工智能导致就业岗位减少的一系列情景进行了比较。事实证明，我们现在预测的范围并不超出常态，至少在未来 20 年内是这样。至于除此之外，谁知道呢？即使有一些非常极端的假设，我们仍然处于历史上看到的转变范围内。

JAMES MANYIKA: Yes, the economy is still made up of a large chunk of those occupations. While some of these will decline, few will disappear entirely and certainly not as quickly as some are predicting. Actually, one of the things that we’ve looked at is where, over the last 200 years, we’ve seen the most massive job declines. For example, we studied what happened to manufacturing in the United States, and the shift from agriculture to industrialization. We looked at 20 different massive job declines in different countries and what happened in each, compared to the range of scenarios for job declines due to automation and AI. It turned out the ranges that we anticipate now are not out of the norm, at least in the next 20 years anyway. Now beyond that, who knows? Even with some very extreme assumptions, we’re still well within the ranges of shifts that we have seen historically.

至少在未来 20 年左右，最大的问题是是否有足够的工作给每个人。正如我们在麦肯锡全球研究院讨论的那样，除非我们做出非常极端的假设，否则我们得出的结论是，每个人都有足够的工作。我们必须问自己的另一个重要问题是，在那些正在衰落的职业和那些正在增长的职业之间，我们将看到多大的转变？我们将看到从一个职业到另一个职业的转变程度有多大？工作场所需要在多大程度上进行调整和适应机器对人的补充，而不是人们失去工作？

The big question, at least in the next 20 or so years, is whether there will be enough work for everybody. As we discussed, at MGI we conclude there will be enough work for everybody, unless we get to those very extreme assumptions. The other important question we must ask ourselves is how big are the scale of transitions that we’ll see between those occupations that are declining, and those occupations that are be growing? What level of movement will we see from one occupation to another, and how much will the workplace need to adjust and adapt to machines complementing people as opposed to people losing their jobs?

根据我们的研究，我们并不确信，按照我们目前的进程和速度，我们在技能、教育和在职培训方面已经做好了应对这些转变的准备。事实上，我们更担心的是转变问题，而不是“是否有足够的工作？”的问题。

Based on our research, we’re not convinced that on our current course and speed we’re well set up to manage those transitions in terms of skilling, educating, and on-the-job training. We actually worry more about that question of transition than about the “Will there be enough work?” question.

马丁·福特：那么，未来真的有可能出现严重的技能不匹配现象吗？

MARTIN FORD: So, there really is the potential for a severe skill mismatch scenario going forward?

詹姆斯·马尼卡：是的，技能不匹配是一个很大的问题。行业和职业的变化，人们必须从一个职业转到另一个职业，适应更高或更低的技能，或者只是不同的技能。

JAMES MANYIKA: Yes, skill mismatches is a big one. Sectoral and occupation changes, where people have to move from one occupation to another, and adapt to higher or lower skill, or just different skills.

当你从行业和地理位置的角度来看待转型时，比如在美国，会有足够的工作，但当你进一步深入研究这些工作的可能地点时，就会发现地理位置不匹配的可能性，有些地方看起来比其他地方更糟糕。这类转型相当重大，我们是否做好了准备还不太清楚。

When you look at the transition in terms of sectoral and geographic locational questions, say in the United States, there will be enough work, but then you go down to the next level to look at the likely locations for that work, and you see the potential for geographic locational mismatches, where some places look like they’ll be more in a hole than other places. These kinds of transitions are quite substantial, and it’s not quite clear if we’re ready for them.

对工资的影响是另一个重要问题。如果你看看可能的职业变化，你会发现很多可能衰落的职业往往是中等工资的职业，比如会计师。许多高薪职业都涉及某种形式的数据分析。它们还涉及高度结构化的环境中的体力劳动，比如制造业。因此，许多即将衰落的职业就处于工资范围内。而许多将增长的职业——比如我们刚才谈到的护理工作——在目前的工资结构下，这些职业的工资并不高。这些职业组合的变化可能会导致严重的工资问题。我们需要改变这些工资动态运作的市场机制，或者开发一些其他机制来塑造这些工资的结构。

The impact on wages is another important question. If you look at the likely occupational shifts, so many of the occupations that are likely to decline have tended to be the middle-wage occupations like accountants. Many well-paying occupations have involved data analysis in one form or another. They have also involved physical work in highly structured environments, like manufacturing. And so that’s where many of the occupations that are going to decline sit on the wage spectrum. Whereas many of the occupations that are going to grow—like the care work we just talked about—are occupations that, at today’s current wage structures, don’t pay as well. These occupational mix shifts will likely cause a serious wage issue. We will need to either change the market mechanisms for how these wage dynamics work or develop some other mechanisms that shape the way these wages are structured.

担心工资问题的另一个原因来自于对我们许多技术专家迄今为止所说的叙述的更深入研究。当我们说“不，别担心。我们不会取代工作，机器将补充人们的工作”时，我认为这是对的，我们自己的麦肯锡全球研究院分析表明，60% 的职业只有大约三分之一的活动将由机器自动化，这意味着人们将与机器一起工作。

The other reason to worry about the wage question comes from a deeper examination of the narrative that many of us as technologists have been saying so far. When we say, “No, don’t worry about it. We’re not going to replace jobs, machines are going to complement what people do,” I think this is true, our own MGI analysis suggests that 60% of occupations will only have about a third of their activities automated by machines, which means people will be working alongside machines.

但是如果我们从工资的角度来审视这一现象，结果就不那么明确了，因为我们知道，当人类与机器相辅相成时，可以产生一系列的结果。例如，我们知道，如果一名高技能工人与一台机器相辅相成，机器做它最擅长的事情，而人类仍然在做高附加值的工作来补充机器，那就太好了。这项工作的工资可能会上涨，生产率也会提高，一切都会非常顺利，这是一个很好的结果。

But if we examine this phenomenon with wages in mind, it’s not so clear-cut because we know that when people are complemented by machines, you can have a range of outcomes. We know that, for example, if a highly skilled worker is complemented by a machine, and the machine does what it does best, and the human is still doing highly value-added work to complement the machine, that’s great. The wages for that work are probably going to go up, productivity will go up and it’ll all work out wonderfully well all round, which is a great outcome.

然而，我们也可能看到另一个极端，即如果人被机器所取代——即使机器只完成了 30% 的工作，但机器却完成了所有增值工作——那么人类剩下的工作就不再那么复杂或不再那么熟练了。这可能会导致工资下降，因为现在有更多人可以完成那些以前需要专业技能或认证的任务。这意味着，将机器引入该职业可能会给该职业的工资带来压力。

However, we could also have the other end of the spectrum, where if the person’s being complemented by a machine—even if the machine is only 30% of the work, but the machine is doing all the value-added portion of that work—then what’s left over for the human being is deskilled or less complex. That can lead to lower wages because now many more people can do those tasks that previously required specialized skills, or required a certification. That means that what you’ve done by introducing machines into that occupation could potentially put pressure on wages in that occupation.

这种补充工作的想法具有广泛的潜在结果，而我们往往只庆祝结果范围的一端，而很少谈论另一端，即技能低下的一端。顺便说一句，这也增加了不断重新学习技能的挑战，因为人们与不断发展和能力越来越强的机器一起工作。

This idea of complementing work has this wide range of potential outcomes, and we tend just to celebrate the one end of the result spectrum, and not talk as much about the other, deskilled, end of the spectrum. This by the way also increases the challenge of reskilling on an ongoing basis as people work alongside ever evolving and increasingly capable machines.

马丁·福特：GPS 对伦敦出租车司机的影响就是一个很好的例子。

MARTIN FORD: A good example of that is the impact of GPS on London taxi drivers.

詹姆斯·马尼卡：是的，这是一个很好的例子，说明劳动力提供的限制部分实际上是伦敦出租车司机头脑中所有街道和捷径的“知识”。当你因为 GPS 系统而贬低这项技能时，剩下的就只是驾驶了，更多的人可以开车带你从 A 地到 B 地。

JAMES MANYIKA: Yes, that’s a great example of where the labor-supplied limiting portion was really “the Knowledge” of all the streets and shortcuts in the minds of the London taxi drivers. When you devalue that skill because of GPS systems, what’s left over is just the driving, and many more people can drive and get you from A to B.

另一个例子是，以呼叫中心接线员为例，这是一种旧的去技能化形式。过去，呼叫中心人员实际上必须知道他们在技术层面上经常谈论什么，才能为您提供帮助。然而，如今，组织将这些知识嵌入到他们阅读的脚本中。剩下的大部分只是能够阅读脚本的人。他们实际上并不需要知道技术细节，至少不像以前那么多；他们只需要能够理解和阅读脚本，除非他们遇到真正的极端情况，在这种情况下他们可以升级为深度专家。

Another example here, in an old form of deskilling, is to think about call center operators. It used to be that your call center person actually had to know what they were talking about often at a technical level in order to be helpful to you. Today, however, organizations embedded that knowledge into the script that they read. What’s left over for the most part is just someone who can read a script. They don’t really need to know the technical details, at least not as much as before; they just need to be able to follow and read the script, unless they get to a real corner case, where they can escalate to a deep expert.

有很多服务工作和服务技术人员工作的例子，无论是通过呼叫中心，还是亲自到现场维修，其中一些工作正在经历大量的去技能化——因为知识嵌入在技术、脚本或其他方式中，以封装解决问题所需的知识。最后，剩下的是一些更加去技能化的东西。

There are many examples of service work and service technician work, whether it’s through the call center, or even people physically showing up to done on-site repairs, where some portions of that work are going through this massive deskilling—because the knowledge is embedded in either technology, or scripts, or some other way to encapsulate the knowledge required to solve the problem. In the end, what’s left over is something much more deskilled.

马丁·福特：那么，听起来总体而言，你更关心工资受到的影响而不是彻底的失业？

MARTIN FORD: So, it sounds like overall, you’re more concerned about the impact on wages than outright unemployment?

詹姆斯·马尼卡：当然，我们总是担心失业问题，因为总有这种极端情况发生，就就业而言，这会导致我们彻底完蛋。但我更担心的是这些劳动力转型问题，比如技能转型、职业转型，以及我们将如何帮助人们度过这些转型期。

JAMES MANYIKA: Of course you always worry about unemployment, because you can always have this corner-case scenario that could play out, which results in a game over for us as far as employment is concerned. But I worry more about these workforce transition issues, such as skills shifts, occupational shifts and how will we support people through these transitions.

我还担心工资效应，除非我们改变劳动力市场对工作的评价方式。从某种意义上说，这个问题已经存在了一段时间。我们都说我们重视照顾孩子的人，我们重视教师；但我们从未在这些职业的工资结构中充分反映这一点，而且这种差异可能很快就会变得更大，因为许多可能增长的职业都将是这样的。

I also worry about the wage effects, unless we evolve how we value work in our labor markets. In a sense this problem has been around for a while. We all say that we value people who look after our children, and we value teachers; but we’ve never quite reflected that in the wage structure for those occupations, and this discrepancy could soon get much bigger, because many of the occupations that are likely to grow are going to look like that.

马丁·福特：正如您之前提到的，这会反过来引发消费需求问题，而这本身就会抑制生产力和增长。

MARTIN FORD: As you noted earlier, that can feed back into the consumer-demand problem, which in itself dampens down productivity and growth.

詹姆斯·马尼卡：当然。这会形成恶性循环，进一步损害工作需求。我们需要迅速采取行动。再培训和在职培训之所以非常重要，首先是因为这些技能变化非常快，人们需要非常迅速地适应。

JAMES MANYIKA: Absolutely. That would create a vicious cycle that further hurts demand for work. And we need to move quickly. The reason why the reskilling and on-the-job training portions are a really important thing is, first of all, because those skills are changing pretty rapidly, and people are going to need to adapt pretty rapidly.

我们已经遇到了问题。我们在研究中指出，如果你看看大多数发达经济体在职培训上花费了多少，你会发现过去 20 到 30 年来，在职培训水平一直在下降。考虑到在职培训在不久的将来会成为一件大事，这是一个真正的问题。

We already have a problem. We have pointed this out in our research that if you look across most advanced economies at how much these countries spend on on-the-job training, the level of on-the-job training has been declining in the last 20 to 30 years. Given that on-the-job training is going to be a big deal in near the future, this is a real issue.

您还可以查看的另一个指标是通常所说的“积极的劳动力市场支持”。这些与在职培训无关，而是在工人被迫从一种职业转向另一种职业时为他们提供的支持。我认为这是我们在上一轮全球化中搞砸的事情之一。

The other measure you can also look at is what is typically called “active labor-market supports.” These are things that are separate from on-the-job training and are instead the kind of support you provide workers when they’re being displaced, as they transition from one occupation to another. This is one of the things I think we screwed up in the last round of globalization.

谈到全球化，人们可以整天争论全球化对生产力、经济增长、消费者选择和产品有何好处。这些都是正确的，但如果你从工人的角度来看待全球化问题，那么它就有问题了。没有有效地为流离失所的工人提供支持。尽管我们知道全球化的痛苦高度局限于特定的地方和部门，但它们仍然足够严重，并且确实影响了许多真实的人和社区。如果你和你的 9 个朋友在 2000 年在美国从事服装制造业，十年后只有 3 个工作岗位仍然存在，如果你和你的 9 个朋友在纺织厂工作，情况也是如此。以密西西比州韦伯斯特县为例，由于服装制造业的遭遇，该县三分之一的工作岗位流失，而服装制造业是该社区的重要组成部分。我们可以说，这在总体上可能会奏效，但如果你就是这些受灾特别严重的社区的工人之一，那就不太令人欣慰了。

With globalization, one can argue all day along about how globalization was great for productivity, economic growth, for consumer choice, and for products. All true, except when you look at the question of globalization through the worker lens; then it’s problematic. The thing that didn’t happen effectively was providing support for the workers who were displaced. Even though we know the pain of globalization was highly localized in specific places and sectors, they were still significant enough and really affected many real people and communities. If you and your 9 friends worked in apparel manufacturing in the US in 2000, a decade later only 3 of those jobs still exist, and the same is true if you and your 9 friends worked in a textile mill. Take Webster County in Mississippi where one third of jobs were lost due to what happened to apparel manufacturing, which was a major part of that community. We can say this will probably work out at an overall level, but that isn’t very comforting if you’re one of the workers in these particularly hard-hit communities.

如果我们说我们需要支持那些已经和即将因工作转型而失业的工人，他们需要从一份工作转到另一份工作，从一种职业转到另一种职业，从一种技能转到另一种技能，那么我们就落后了。因此，工人转型挑战确实是一个大问题。

If we say that we’re going to need to support both workers who have been, and those who are going to be, dislocated through these work transitions and will need to go from one job to another, or one occupation to another, or one skill-set to another, then we’re starting from behind. So, the worker transition challenges are a really big deal.

马丁·福特：您说的是，我们需要支持工人，无论他们是失业还是转型。您认为全民基本收入可能是实现这一目标的好办法吗？

MARTIN FORD: You’re making the point that we’re going to need to support workers, whether they’re unemployed or they’re transitioning. Do you think a universal basic income is potentially a good idea for doing that?

詹姆斯·马尼卡：我对全民基本收入的想法有以下几点矛盾。我喜欢我们讨论这个问题，因为这承认了我们可能存在工资和收入问题，而且这在世界范围内引发了一场争论。

JAMES MANYIKA: I’m conflicted about the idea of universal basic income in the following sense. I like the fact that we’re discussing it, because it’s an acknowledgment that we may have a wage and income issue, and it’s provoking a debate in the world.

我认为，它忽视了工作所发挥的更广泛的作用。工作是一件复杂的事情，因为工作不仅能提供收入，还能做很多其他的事情。它提供了意义、尊严、自尊、目的、社区和社会影响等等。建立一个基于 UBI 的社会，虽然可以解决工资问题，但不一定能解决工作带来的其他方面的问题。而且，我认为我们应该记住，还有很多工作要做。

My issue with it is that I think it misses the wider role that work plays. Work is a complicated thing because while work provides income, it also does a whole bunch of other stuff. It provides meaning, dignity, self-respect, purpose, community and social effects, and more. By going to a UBI-based society, while that may solve the wage question, it won’t necessarily solve these other aspects of what work brings. And, I think we should remember that there will still be lots of work to be done.

令我印象深刻、非常有趣的一句话来自林登·约翰逊总统的“技术、自动化和经济进步”蓝丝带委员会，其中鲍勃·索洛也在其中。该报告的结论之一是：“基本事实是，技术消除的是工作岗位，而不是工作。”

One of the quotes that really sticks with me and I find quite fascinating is from President Lyndon B. Johnson’s Blue-Ribbon Commission on “Technology, Automation, and Economic Progress,” which incidentally included Bob Solow. One of the report’s conclusions is that “The basic fact is that technology eliminates jobs, not work.”

马丁·福特：总有工作要做，但劳动力市场可能不重视。

MARTIN FORD: There’s always work to be done, but it might not be valued by the labor market.

詹姆斯·马尼卡：它并不总是出现在我们的劳动力市场中。想想看，在大多数社会中，护理工作往往由女性承担，而且往往没有报酬。我们如何在劳动力市场和工资收入讨论中反映护理工作的价值？这项工作将存在。只是它是否有偿工作，或者是否被认可为工作，并以这种方式获得补偿。

JAMES MANYIKA: It doesn’t always show up in our labor markets. Just think about care work, which in most societies tends to be done by women and is often unpaid. How do we reflect the value of that care work in our labor markets and discussions on wages and incomes? The work will be there. It’s just whether it’s paid work, or recognized as work, and compensated in that way.

我喜欢 UBI 引发关于工资和收入的讨论，但我不确定它是否能像其他事情一样有效地解决工作问题。我更愿意考虑有条件转移之类的概念，或者其他一些方法，以确保我们将工资与某种反映主动性、目的、尊严和其他重要因素的活动联系起来。这些关于目的、意义和尊严的问题最终可能就是我们定义的东西。

I like the fact that UBI is provoking the conversation about wages and income, but I’m not sure it solves the work question as effectively as other things might do. I prefer to consider concepts like conditional transfers, or some other way to make sure that we are linking wages to some kind of activity that reflects initiative, purpose, dignity, and other important factors. These questions of purpose, meaning and dignity may in the end be what defines us.

詹姆斯·马尼卡 是麦肯锡公司高级合伙人兼麦肯锡全球研究院 (MGI) 主席。詹姆斯还是​​麦肯锡董事会成员。詹姆斯在硅谷工作了 20 多年，曾与许多世界领先科技公司的首席执行官和创始人就各种问题进行过合作。在麦肯锡全球研究院，詹姆斯领导了技术、数字经济以及增长、生产力和全球化方面的研究。他出版了一本关于人工智能和机器人的书，另一本关于全球经济趋势的书，以及在商业媒体和学术期刊上发表的众多文章和报告。

JAMES MANYIKA is a senior partner at McKinsey & Company and chairman of the McKinsey Global Institute (MGI). James also serves on McKinsey’s Board of Directors. Based in Silicon Valley for over 20 years, James has worked with the chief executives and founders of many of the world’s leading technology companies on a variety of issues. At MGI, James has led research on technology, the digital economy, as well as growth, productivity, and globalization. He has published a book on AI and robotics, another on global economic trends as well as numerous articles and reports that have appeared in business media and academic journals.

詹姆斯被奥巴马总统任命为白宫全球发展委员会副主席（2012-16 年），并被商务部长任命为美国商务部数字经济顾问委员会和国家创新顾问委员会成员。他还是外交关系委员会、约翰·麦克阿瑟基金会、休利特基金会和马克尔基金会的董事会成员。

James was appointed by President Obama as vice chair of the Global Development Council at the White House (2012-16) and by Commerce Secretaries to the US Commerce Department’s Digital Economy Board of Advisors and the National Innovation Advisory Board. He serves on the boards of the Council on Foreign Relations, John D. and Catherine T. MacArthur Foundation, Hewlett Foundation, and Markle Foundation.

他还担任牛津互联网研究所、麻省理工学院数字经济计划等学术顾问委员会成员。他是斯坦福大学百年人工智能研究常务委员会委员、AIIndex.org 团队成员和 DeepMind 研究员。

He also serves on academic advisory boards including the Oxford Internet Institute, MIT’s Initiative on the Digital Economy. He is on the standing committee for the Stanford-based 100 Year Study on Artificial Intelligence, a member of the AIIndex.org team, and a fellow at DeepMind.

James 曾任牛津大学工程系教授、编程研究小组和机器人研究实验室成员、牛津大学贝利奥尔学院研究员、美国宇航局喷气推进实验室客座科学家和麻省理工学院交流研究员。James 是罗德学者，在牛津大学获得机器人、数学和计算机科学博士学位、理学硕士和文学硕士学位，并以英美学者身份在津巴布韦大学获得电气工程学士学位。

James was on the engineering faculty at Oxford University and a member of the Programming Research Group and the Robotics Research Lab, a fellow of Balliol College, Oxford, a visiting scientist at NASA Jet Propulsion Labs, and a faculty exchange fellow at MIT. A Rhodes Scholar, James received his DPhil, MSc, and MA from Oxford in Robotics, Mathematics, and Computer Science, and a BSc in electrical engineering from University of Zimbabwe as an Anglo-American scholar.

GEOMETRIC INTELLIGENCE（被 UBER 收购）创始人兼首席执行官纽约大学心理学和神经科学教授

FOUNDER AND CEO, GEOMETRIC INTELLIGENCE (ACQUIRED BY UBER) PROFESSOR OF PSYCHOLOGY AND NEURAL SCIENCE, NYU

加里·马库斯是 Uber 收购的机器学习公司 Geometric Intelligence 的创始人兼首席执行官，也是纽约大学心理学和神经科学教授，以及多部书籍的作者和编辑，包括《大脑的未来》和畅销书《吉他零》。加里的大部分研究都集中在了解儿童如何学习和吸收语言。他目前的工作是研究人类思维的洞察力如何为人工智能领域提供信息。

Gary Marcus was Founder and CEO of Geometric Intelligence, a machine learning company acquired by Uber, and is a professor of psychology and neural science at New York University, as well as the author and editor of several books, including The Future of the Brain and the bestseller Guitar Zero. Much of Gary’s research has focused on understanding how children learn and assimilate language. His current work is on how insights from the human mind can inform the field of artificial intelligence.

马丁·福特：克鲁格，你写了一本书，讲述大脑是如何成为一个不完美的器官；那么，你想必不认为 AGI 的途径就是试图完美地复制人类的大脑吧？

MARTIN FORD: You wrote a book, Kluge, about how the brain is an imperfect organ; presumably, then, you don’t think the route to AGI is to try to perfectly copy the human brain?

加里·马库斯：不，我们不需要复制人类大脑及其所有低效之处。有些事情人类做得比现在的机器好得多，你想从中学习，但还有很多事情你不想复制。

GARY MARCUS: No, we don’t need to replicate the human brain and all of its inefficiencies. There are some things that people do much better than current machines, and you want to learn from those, but there are lots of things you don’t want to copy.

我不确定 AGI 系统看起来会有多像人类。但是，人类是目前我们所知道的唯一能够根据非常广泛的数据进行推断和规划并以非常有效的方式进行讨论的系统，因此研究人类如何做到这一点是值得的。

I’m not committed to how much like a person an AGI system will look. However, humans are currently the only system that we know of that can make inferences and plans over very broad ranges of data and discuss them in a very efficient way, so it pays to look into how people are doing that.

我写的第一本书名为《代数思维》 ，出版于 2001 年，书中将神经网络与人类进行了比较。我探索了如何才能让神经网络变得更好，我认为这些论点在今天仍然非常有意义。

The first book that I wrote, published in 2001, was titled The Algebraic Mind, and it compared neural networks with humans. I explored what it would take to make neural networks better, and I think those arguments are still very relevant today.

我写的下一本书名为《心灵的诞生》，内容是了解基因如何构建我们心灵中的先天结构。它源自诺姆·乔姆斯基和史蒂芬·平克的传统，他们相信心灵中存在着重要的东西。在书中，我试图从分子生物学和发育神经科学的角度来理解先天性可能意味着什么。同样，我认为书中的观点在今天仍然具有现实意义。

The next book I wrote was called The Birth of the Mind, and was about understanding how genes can build the innate structures in our mind. It comes from the Noam Chomsky and Steven Pinker tradition of believing that there are important things built into the mind. In the book, I tried to understand what innateness might mean in terms of molecular biology and developmental neuroscience. Again, I think the ideas there are quite relevant today.

2008 年，我出版了《克鲁格：人类思维的随机进化》。有些人可能不知道，“克鲁格”是老工程师对问题的笨拙解决方案的称呼。在那本书中，我认为人类思维在很多方面实际上就是这样的。我研究了关于人类是否是最佳状态的讨论——我认为人类显然不是——并试图从进化的角度理解为什么我们不是最佳状态。

In 2008 I published Kluge: The Haphazard Evolution of the Human Mind. For those who may not know, “kluge” is an old engineer’s term for a clumsy solution to a problem. In that book, I argued that in many ways the human mind was actually something like that. I examined discussions about whether humans are optimal—to which I think they’re clearly not—and tried to understand from an evolutionary perspective why we’re not optimal.

马丁·福特：这是因为进化必须从现有的框架开始，然后在此基础上进行构建，对吗？它不可能回到过去，从头开始重新设计一切。

MARTIN FORD: That’s because evolution has to work from an existing framework and build from there, right? It can’t go back and redesign everything from scratch.

加里·马库斯：没错。这本书的很多内容都是关于我们的记忆结构，以及它与其他系统的比较。例如，当你将我们的听觉系统与理论上可能的听觉系统进行比较时，我们会发现我们非常接近最佳状态。如果你将我们的眼睛与理论上的最佳状态进行比较，我们再次接近最佳状态——在适当的条件下，你可以看到一个光子，这真是太神奇了。然而，我们的记忆力并不是最佳的。

GARY MARCUS: Exactly. A lot of the book was about our memory structure, and how that compares to other systems. For example, when you compare our auditory systems to what’s theoretically possible, we come very close to optimal. If you compare our eyes to the theoretical optimum, we’re again close—given the right conditions, you can see a single photon of light, and that’s amazing. Our memory, however, is not optimal.

你可以非常快速地将莎士比亚的全部作品上传到计算机，或者实际上，将有史以来的大部分作品上传到计算机，计算机不会忘记其中的任何内容。我们的记忆在容量或存储的记忆稳定性方面远未达到理论上的最佳状态。我们的记忆往往会随着时间的推移而变得模糊。如果你每天都把车停在同一个停车位，你就记不住今天把车停在哪里了，因为你无法将今天的记忆与昨天的记忆区分开来。计算机永远不会遇到这个问题。

You could very quickly upload the complete works of Shakespeare to a computer, or in fact, most of what’s been written ever, and a computer won’t forget any of it. Our memories are nowhere near theoretically optimal in terms of their capacity or in terms of the stability of the memory that you store. Our memories tend to blur together over time. If you park in the same space every day you can’t remember where you parked today, because you can’t keep today’s memory distinct from yesterday’s memory. A computer would never have trouble with that.

我在书中提出的论点是，我们可以研究并理解人类为何会有如此糟糕的记忆，从我们的祖先需要从他们的记忆中得到什么的角度来说。它主要是广泛的统计总结，例如：“山上的食物比山下多。”我不需要记住我从哪一天获得了这些记忆痕迹，我只需要知道山上比山下更肥沃的总体趋势。

The argument I made in the book was that we could examine and understand why humanity had such crummy memories, in terms of what our ancestors needed from their memory. It was mostly broad statistical summaries like: “there’s more food up the mountain than down the mountain.” I don’t need to remember what individual days I derived those memory traces from, I just need the general trend that it’s more fertile up the mountain as opposed to down the mountain.

脊椎动物进化出了这种记忆，而不是计算机使用的那种，后者是一种位置可寻址的记忆，其中每个位置都分配给一个特定的稳定功能。这就是让你能够在计算机上存储几乎无限的信息而不会出现混淆问题的原因。人类在进化链中走的是一条不同的道路，从基因数量来看，我们需要改变这些基因，才能从头开始重建一个围绕位置可寻址记忆的系统，这将非常昂贵。

Vertebrates evolved that kind of memory—instead of what computers use, which is a location-addressable memory where every single location in the memory is assigned to a particular stable function. That’s what allows you to store essentially infinite information on a computer without having the problem of blurring things together. Humans went down a different path in the evolutionary chain, and it would be very costly in terms of the number of genes that we would need to change in order to just rebuild the system from scratch around location-addressable memory.

实际上，构建混合系统是可能的。谷歌就是一个混合系统，因为它的底层是位置可寻址内存，顶层是提示可寻址内存，而这正是我们所拥有的。这是一个好得多的系统。谷歌可以像我们一样接受提醒提示，但它有一张所有事物所在位置的总图，因此它提供了正确的答案，而不是任意扭曲答案。

It’s actually possible to build hybrids. Google is a hybrid, as it has location-addressable memory underneath and then cue-addressable memory, which is what we have, on top. That’s a much better system. Google can take reminder cues as we can, but then it has a master map of where everything is, so it serves up the right answer instead of arbitrarily distorting the answer.

马丁·福特：您能详细解释一下吗？

MARTIN FORD: Could you explain that in more detail?

加里·马库斯：线索寻址记忆是指记忆由其他因素触发或辅助。这种记忆有各种疯狂的版本，比如姿势依赖性记忆。如果你站着学到了一些东西，那么如果你站着回忆，你会比躺着回忆时记得更牢。最臭名昭著的是状态依赖性记忆。例如，如果你在吸毒后复习考试，那么在考试时吸毒可能对你更有好处。我不建议这样做……关键是你的姿势和周围的线索会影响你的记忆。

GARY MARCUS: Cue-addressable memory is where memories are triggered or aided by other factors. There are crazy versions of this like posture-dependent memory. This is where if you learned something standing up then you’ll remember it better if you try to recall it standing up than if you’re lying down. The most notorious one is state-dependent memory. For example, if you study for an exam while you’re stoned, you might actually be better off being stoned when you take the exam. I don’t suggest doing that…the point is that the state and the cues around you influence what you remember.

另一方面，你不能说“我想要内存位置 317”或“我在 1997 年 3 月 17 日学到的东西”。作为人类，你无法像计算机那样提取东西。计算机的这些索引实际上就像一组邮政信箱，而放在 972 号信箱里的东西会无限期地留在那里，除非你故意篡改它。

On the other hand, you can’t say, “I want memory location 317” or “the thing I learned on March 17, 1997.” As a human, you can’t pull things out the way a computer could. A computer has these indexes that are actually like a set of post-office boxes, and what is put in box number 972 stays there indefinitely, unless you deliberately tamper with it.

我们的大脑似乎对此没有把握。大脑没有内部寻址系统来知道单个记忆存储在哪里。相反，大脑似乎在做一些更像拍卖的事情。它会说，“有什么东西可以告诉我在阳光明媚的日子里我应该在车里做什么？”你得到的是一组相关的记忆，但至少在意识上不知道它们在大脑中的物理存储位置。

It doesn’t even appear that our brain has a handle on this. The brain does not have an internal addressing system to know where individual memories are stored. Instead, it seems like the brain does something more like an auction. It says, “Is there anything out there that can give me information about what I should be doing in a car on a sunny day?” What you get back is a set of relevant memories without knowing, at least consciously, where they are physically stored in the brain.

问题是，它们有时会混杂在一起，例如，这会导致目击证人的证词出现问题。你实际上无法将某一特定时刻发生的事情与你后来想到的事情、你在电视上看到的事情或在报纸上读到的事情区分开来。所有这些事情都混杂在一起，因为它们没有被清晰地存储起来。

The problem is sometimes they blur together, and that leads for example, to problems with eyewitness testimonies. You can’t actually keep the state of what happened at a particular moment, separate from what you thought about later, or what you saw on television or read in the newspaper. All these things blur together because they’re not distinctly stored.

马丁·福特：这很有趣。

MARTIN FORD: That’s interesting.

加里·马库斯：我的书《克鲁格》的第一个核心观点是，记忆基本上有两种，而人类只能使用不太有用的一种。我进一步指出，一旦我们在进化史上拥有了这种记忆，你就不可能从头开始，所以你只能在此基础上进行构建。这就像斯蒂芬·杰伊·古尔德关于熊猫拇指的著名论点。

GARY MARCUS: The first central claim of my book Kluge was that there are basically two kinds of memory and that humans got stuck with the one that’s less useful. I further argued that once we have that in our evolutionary history, it becomes astonishingly unlikely that you’re going to start from scratch, so you just build on top of that. This is like Stephen Jay Gould’s famous arguments about the panda’s thumb.

一旦你有了这种记忆，其他的东西也会随之而来，比如确认偏差。确认偏差是指你对与你的理论一致的事实的记忆比对与你的理论不一致的事实的记忆要好。计算机不需要这样做。计算机可以搜索与邮政编码匹配的所有内容或与邮政编码不匹配的所有内容。它可以使用 NOT 运算符。使用计算机，我可以搜索邮政编码中的所有男性和 40 岁以上的人，或者同样搜索所有不符合这些条件的人。人类大脑使用线索可寻址内存，只能在数据中搜索匹配项。其他一切都要困难得多。

Once you have that kind of memory, other things come with it, such as confirmation bias. Confirmation bias is where you remember facts that are consistent with your theory better than facts that are inconsistent with your theory. A computer doesn’t need to do that. A computer can search for everything that matches a zip code or everything that doesn’t match a zip code. It can use NOT operators. Using a computer, I can search for everybody that is male and over 40 in my zip code, or equally everybody that doesn’t match those criteria. The human brain, using cue-addressable memory, can only search for matches within data. Everything else is much harder.

如果我有一个理论，那么我就能找到符合我理论的材料，但任何不符合的材料就不那么容易被想到。我无法系统地搜索它。这就是确认偏差。

If I have a theory then I can find material that matches my theory, but anything that doesn’t match doesn’t come to mind as easily. I can’t systematically search for it. That’s confirmation bias.

另一个例子是聚焦幻觉，我以两种顺序之一问你两个问题。我要么问你对婚姻有多幸福，然后问你对生活有多幸福，要么以另一种顺序问你。如果我先问你对婚姻有多幸福，这会影响你对生活的看法。你应该能够将这两件事完全分开。

Another example is the focusing illusion, where I ask you two questions in one of two orders. I either ask you how happy are you with your marriage and then how happy are you with your life, or in the other order. If I ask you first how happy you are with your marriage, that influences how you think about your life in general. You should be able to keep the two things completely separate.

马丁·福特：这听起来像丹尼尔·卡尼曼的锚定理论，他谈论的是如何给人们一个随机数，然后这个数字就会影响他们对任何事情的猜测。

MARTIN FORD: That sounds like Daniel Kahneman’s anchoring theory, where he talks about how you can give people a random number, and then that number will influence their guess about anything.

加里·马库斯：是的，这是一个变种。如果我先让你看一美元钞票上的最后三位数字，然后问你《大宪章》签署的时间，你就会记住这三位数字。

GARY MARCUS: Yes, it’s a variation. If I asked you when the Magna Carta was signed, after first asking you to look at the last three digits on a dollar bill, those three digits on the dollar bill anchor your memory.

马丁·福特：你的职业轨迹与人工智能领域的许多其他人截然不同。你早期的工作重点是理解人类语言和儿童学习语言的方式，最近你与他人共同创办了一家初创公司，并帮助建立了 Uber 的人工智能实验室。

MARTIN FORD: Your career trajectory is quite different from a lot of other people in the field of AI. Your early work focused on understanding human language and the way children learn it, and more recently you co-founded a startup company and helped launch Uber’s AI labs.

加里·马库斯：我觉得自己有点像约瑟夫·康拉德（1857-1924），他说波兰语，但用英语写作。虽然他不是英语的母语人士，但他对英语的运作方式有很多见解。同样，我认为自己不是机器学习或人工智能的母语人士，而是一个从认知科学转向人工智能并有新见解的人。

GARY MARCUS: I feel a bit like Joseph Conrad (1857-1924), who spoke Polish but wrote in English. While he wasn’t a native speaker of English, he had a lot of insights into the workings of it. In the same way, I think of myself as not a native speaker of machine learning or AI, but as someone who is coming to AI from the cognitive sciences and has fresh insights.

我在童年时期做过很多计算机编程，也思考过很多关于人工智能的问题，但读研究生时，我对认知科学的兴趣比人工智能更大。在读研究生期间，我跟随认知科学家史蒂芬·平克学习，我们研究了儿童如何学习语言中的过去时，然后利用当时深度学习的前身，即多层和双层感知器，对此进行了研究。

I did a lot of computer programming throughout my childhood and thought a lot about artificial intelligence, but I went to graduate school more interested in the cognitive sciences than artificial intelligence. During my time at graduate school, I studied with the cognitive scientist Steven Pinker, where we looked at how children learn the past tense within a language and then examined that using the precursors to deep learning that we had at the time, namely multi-layer and two-layer perceptrons.

1986 年，David Rumelhart 和 James L. McClelland 发表了一篇题为《并行分布式处理：认知微观结构的探索》的论文，该论文表明神经网络可以学习英语的过去时。Pinker 和我仔细研究了这篇论文，虽然你可以让神经网络过度规则化并像孩子们一样说出“goed”或“breaked”之类的词，但关于他们何时以及如何犯下这些错误的所有事实实际上完全不同。针对这篇论文，我们假设孩子们同时使用了规则和神经网络。

In 1986, David Rumelhart and James L. McClelland published a paper titled Parallel Distributed Processing: explorations in the microstructure of cognition, which showed that a neural network could learn the past tense of English. Pinker and I looked at the paper in some detail, and although it was true that you could get a neural network to overregularize and say things like “goed” or “breaked” like kids do, all of the facts about when and how they made those errors were actually quite different. In response to the paper, we hypothesized that kids use a hybrid of both rules and neural networks.

马丁·福特：您说的是不规则单词的结尾，孩子们有时会错误地把它们变成规则的。

MARTIN FORD: You’re talking about irregular word endings, where kids will sometimes make them regular by mistake.

加里·马库斯：没错，孩子们有时会将不规则动词规则化。我曾经对 11,000 句孩子与父母用过去时动词交谈的话语进行了自动机器驱动分析。在我的研究中，我观察了孩子们何时犯下这些过度规则化错误，并绘制了这些错误的时间过程以及哪些动词更容易出现这些错误。

GARY MARCUS: Right, kids sometimes regularize irregular verbs. I once did an automated machine-driven analysis of 11,000 utterances of kids talking to their parents with past-tense verbs. In my study, I was looking at when kids made these overregularization errors and plotting the time course of the errors and which verbs were more vulnerable to these errors.

我们提出的论点是，孩子们似乎对常规动词有一套规则。例如，他们会添加 -ed，但与此同时，他们也有某种联想记忆，如今你可以将其视为神经网络，来处理不规则动词。这个想法是，如果你在过去时态中将动词“sing”变为“sang”，那么你可能只是在用你的记忆来做这件事。如果你的记忆理解“sing”和“sang”，它会帮助你记住“ring”和“rang”。

The argument that we made was that children seem to have a rule for the regulars. For example, they add -ed, but at the same time, they also had something of an associative memory, which you might think of nowadays as a neural network, to do the irregular verbs. The idea is if you’re inflecting the verb “sing” as “sang” in the past tense, you might be just using your memory for that. If your memory understands “sing” and “sang,” it’ll help you to remember “ring” and “rang.”

但是，如果你对一个听起来与你之前听过的任何单词都不一样的单词进行变位，比如“rouge”，意思是往脸上涂胭脂，那么这个词就不需要听起来与你之前听过的任何单词都一样。你仍然知道要在其后添加 -ed。你可以说，“Diane 昨天给脸上涂了胭脂。”

However, if you inflect a word that doesn’t sound like anything you’ve heard before, like to “rouge,” which would be to apply rouge to your face, then the word doesn’t need to sound like anything you’ve heard before. You’ll still know to add -ed to it. You’d say, “Diane rouged her face yesterday.”

重点是，虽然神经网络在基于相似性的工作方面非常擅长，但在没有相似性但仍能理解规则的工作方面却非常薄弱。那是 1992 年，25 年后，这个基本观点仍然存在。大多数神经网络仍然存在这样的问题：它们非常依赖数据，而且相对于它们所接受的训练，它们不会产生高水平的抽象。

The point of that was that while neural networks are very good at things that work by similarity, they’re very weak at things where you don’t have a similarity but where you still understand the rule. That was 1992, 25 years later and that basic point still persists today. Most neural networks still have the problem that they’re very data-driven, and they don’t induce a high level of abstraction relative to what they’ve been trained on.

神经网络能够捕捉到很多普通情况，但如果你考虑长尾分布，它们的尾部就非常薄弱。下面是一个字幕系统的例子：一个系统可能能够告诉你某张图像是一群孩子在玩飞盘，只是因为有很多这样的图片，但如果你给它看一个贴满贴纸的停车标志，它可能会说这是一台装满食物和饮料的冰箱。这是谷歌字幕的实际结果。这是因为数据库中没有那么多贴满贴纸的停车标志的例子，所以系统表现很差。

Neural networks are able to capture a lot of the garden-variety cases, but if you think about a long-tail distribution, they’re very weak at the tail. Here’s an example from a captioning system: a system might be able to tell you that a particular image is of a group of kids playing frisbee, simply because there are a lot of pictures that are like that, but if you show it a parking sign covered with stickers then it might say it’s a refrigerator filled with food and drinks. That was an actual Google captioning result. That’s because there aren’t that many examples of parking signs covered with stickers in the database, so the system performs miserably.

神经网络无法在某些核心情况之外很好地概括这一关键问题，这是我整个职业生涯中都感兴趣的问题。从我的角度来看，这是机器学习领域尚未真正解决的问题。

That key problem of neural networks not being able to generalize well outside of some core situations has been something that’s interested me for my entire career. From my point of view, it’s something that the machine learning field has still not really come to grips with.

马丁·福特：理解人类语言和学习显然是您研究的支柱之一。我想知道您是否可以深入探讨一下您进行的一些现实生活中的实验？

MARTIN FORD: Understanding human language and learning is clearly one of the pillars of your research. I was wondering if you could delve into some real-life experiments that you’ve undertaken?

加里·马库斯：我从理解人类概括的角度研究这个问题，在研究期间，我针对儿童、成人进行了研究，最终在 1999 年针对婴儿进行了研究，所有研究都表明人类非常擅长抽象。

GARY MARCUS: During my years of studying this from the perspective of understanding human generalization, I did research with children, adults, and ultimately with babies in 1999, all of which pointed to humans being very good at abstraction.

婴儿实验表明，七个月大的婴儿可以听两分钟的人工语法，并能识别出由该语法构成的句子规则。婴儿会听两分钟带有 ABB 语法的句子“la ta ta”和“ga na na”，然后会注意到“wo fe wo”的语法不同（ABA 语法），而“wo fe fe”的语法与他们接受过训练的其他句子相同。

The experiment with babies showed that seven-month-olds could hear two minutes of an artificial grammar and recognize the rules of sentences constructed by that grammar. Babies would listen to sentences like “la ta ta” and “ga na na” for two minutes with A-B-B grammar and would then notice that “wo fe wo” had a different grammar (an A-B-A grammar) as opposed to “wo fe fe” that had the same grammar as the other sentences that they’d been trained on.

这是通过观察时间长短来衡量的。我们发现，如果我们改变语法，观察时间会更长。这项实验确实证实，从生命早期开始，婴儿就有能力识别语言领域中相当深奥的抽象概念。另一位研究人员后来表明，新生儿也能做同样的事情。

This was measured by how long they would look. We found that they would look longer if we changed the grammar. That experiment really nailed that from very early on in life babies have an ability to recognize pretty deep abstractions in the language domain. Another researcher later showed that newborns could do the same thing.

马丁·福特：我知道您对 IBM 的 Watson 非常感兴趣，它让您重新回到了人工智能领域。您能谈谈为什么 Watson 重新点燃了您对人工智能的兴趣吗？

MARTIN FORD: I know that you have a great interest in IBM’s Watson, and that it drew you back into the field of AI. Could you talk about why Watson reignited your interest in artificial intelligence?

加里·马库斯：我曾经对 Watson 持怀疑态度，所以当它在 2011 年首次赢得 Jeopardy 时，我感到很惊讶。作为一名科学家，我训练自己注意自己犯错的地方，我认为自然语言理解对于当代人工智能来说太难了。Watson 不应该能够在 Jeopardy 中击败人类，但它做到了。这让我再次开始思考人工智能

GARY MARCUS: I was skeptical about Watson, so I was surprised when it first won at Jeopardy in 2011. As a scientist, I’ve trained myself to pay attention to the things that I get wrong, and I thought natural language understanding was too hard for a contemporary AI to do. Watson should not be able to beat a human in Jeopardy, and yet it did. That made me start thinking about AI again

我最终发现，Watson 之所以能获胜，是因为它实际上是一个比最初看起来更狭窄的人工智能问题。答案几乎总是如此。在 Watson 的案例中，这是因为 Jeopardy 中大约 95% 的答案最终都是维基百科页面的标题。它不是理解语言、推理语言等等，而是主要从受限集合（即维基百科标题页面）中进行信息检索。实际上，这个问题并不像外行人想象的那么难，但它足够有趣，让我再次思考人工智能。

I eventually figured out that the reason Watson won is because it was actually a narrower AI problem than it first appeared to be. That’s almost always the answer. In Watson’s case it’s because about 95% of the answers in Jeopardy turn out to be the titles of Wikipedia pages. Instead of understanding language, reasoning about it and so forth, it was mostly doing information retrieval from a restricted set, namely the pages that are Wikipedia titles. It was actually not as hard of a problem as it looked like to the untutored eye, but it was interesting enough that it got me to think about AI again.

大约在同一时间，我开始为《纽约客》撰稿，写了很多关于神经科学、语言学、心理学和人工智能的文章。在我的文章中，我试图运用我对认知科学及其相关知识的了解——思维和语言如何运作，儿童的思维如何发展等——以便更好地理解人工智能和人们所犯的错误。

Around the same time, I started writing for The New Yorker, where I was producing a lot of pieces about neuroscience, linguistics, psychology, and also AI. In my pieces, I was trying to use what I knew about cognitive science and everything around that—how the mind and language work, how children’s minds develop, etc.—in order to give me a better understanding of AI and the mistakes people were making.

大约在同一时间，我开始撰写和思考更多关于人工智能的文章。其中一篇是一篇批评雷·库兹韦尔 (Ray Kurzweil) 一本书的文章。另一篇是关于自动驾驶汽车以及当一辆失控的校车向它们疾驰而来时，它们将如何做出决定。还有一篇非常有先见之明的文章批评了深度学习，说我认为，作为一个社区，我们应该把它理解为众多工具中的一种，而不是人工智能的完整解决方案。当我五年前写这篇文章时，我说我不认为深度学习能够做抽象和因果推理之类的事情，如果你仔细观察，你会发现深度学习仍然在努力解决这些问题。

Around the same time, I starting writing and thinking a lot more about AI. One was a critical piece on one of Ray Kurzweil’s books. Another was about self-driving cars and how they would make a decision if an out-of-control school bus were hurtling toward them. Another, very prescient, piece criticized deep learning, saying that I think, as a community, we should understand it as one tool among many, not as a complete solution to AI. When I wrote that piece five years ago, I said that I didn’t think deep learning would be able to do things like abstraction and causal reasoning, and if you look carefully, you’ll see that deep learning is still struggling with exactly that set of problems.

马丁·福特：我们来谈谈你在 2014 年创办的公司 Geometric Intelligence。我知道这家公司最终被 Uber 收购，不久之后你就加入了 Uber，并成为该公司人工智能实验室的负责人。你能向我们介绍一下这段经历吗？

MARTIN FORD: Let’s talk about the company you started in 2014, Geometric Intelligence. I know that was eventually bought by Uber, shortly after which you moved to Uber and became the head of their AI labs. Can you take us through that journey?

GARY MARCUS：早在 2014 年 1 月，我就意识到，与其写有关 AI 的文章，我其实应该尝试创办一家自己的公司。我招募了一些优秀的人才，包括我的朋友 Zoubin Ghahramani，他是世界上最好的机器学习专家之一。接下来的几年里，我经营着一家机器学习公司。我学到了很多关于机器学习的知识，并提出了一些如何更好地推广的想法。这成为了我们公司的核心知识产权。我们花了很多时间尝试让算法更有效地从数据中学习。

GARY MARCUS: Back in January 2014 it occurred to me that instead of writing about AI I should actually try to start a company of my own. I recruited some great people, including my friend Zoubin Ghahramani, who is one of the best machine learning people in the world, and I spent the next couple of years running a machine learning company. I learned a lot about machine learning and we built on some ideas of how to generalize better. That became our company’s core intellectual property. We spent a lot of time trying to make algorithms learn more efficiently from data.

深度学习对于解决问题所需的数据量要求极高。这在虚拟世界中效果很好，比如围棋，但在现实世界中效果就不是那么好了，因为数据通常很昂贵或难以获得。我们花了很多时间试图在这个领域做得更好，并取得了一些不错的成果。例如，我们可以用深度学习一半的数据量来完成任意任务，比如 MNIST 字符识别任务。

Deep learning is incredibly greedy in terms of the amount of data that it needs in order to solve a problem. That works well in artificial worlds, such as the game of Go, but it doesn’t work that well over in the real world, where data is often expensive or difficult to obtain. We spent a lot of our time trying to do better in that area and had some nice results. For example, we could learn arbitrary tasks like the MNIST character recognition task with half as much data as deep learning.

消息传开后，我们最终于 2016 年 12 月将公司出售给了 Uber。整个过程让我学到了很多关于机器学习的知识，包括它的优点和缺点。我在 Uber 工作过一段时间，帮助启动 Uber AI 实验室，然后离开了。从那时起，我一直在研究如何将 AI 与医学结合起来，也对机器人技术进行了很多思考。

Word got around, and eventually, we sold to Uber in December 2016. This entire process taught me quite a bit about machine learning, including its strengths and weaknesses. I worked briefly at Uber, helping with the launch of Uber AI labs, and then moved on. Since then, I’ve been researching into how AI and medicine can be combined, and also thinking a lot about robotics.

2018 年 1 月，我写了两篇论文（https://arxiv.org/abs/1801.00631）以及 Medium 上的几篇文章。其中一篇是关于深度学习的，尽管深度学习非常流行，也是目前我们最好的人工智能工具，但它不会让我们实现 AGI（通用人工智能）。第二篇文章是关于先天性的，至少在生物学中，系统从许多固有结构开始，无论是心脏、肾脏还是大脑。大脑的初始结构对于我们如何理解世界非常重要。

In January of 2018, I wrote two papers (https://arxiv.org/abs/1801.00631) as well as a couple of pieces on Medium. One strand of that was about deep learning and how although it’s very popular and our best tool for AI at the moment, it’s not going to get us to AGI (artificial general intelligence). The second piece was about innateness, saying that, at least in biology, systems start with a lot of inherent structure, whether you’re talking about the heart, the kidney, or the brain. The brain’s initial structure is important for how we go about understanding the world.

人们谈论先天与后天，但实际上先天与后天共同作用。先天构建了学习机制，使我们能够以有趣的方式利用我们的经验。

People talk about Nature versus Nurture, but it’s really Nature and Nurture working together. Nature is what constructs the learning mechanisms that allow us to make use of our experience in interesting ways.

马丁·福特：婴儿实验证明了这一点。他们还没有来得及学习任何东西，但他们仍然可以做一些基本的事情，比如识别面孔。

MARTIN FORD: That’s something that’s demonstrated by experiments with very young babies. They haven’t had time to learn anything, but they can still do essential things like recognize faces.

加里·马库斯：没错。我对八个月大婴儿的研究也证实了这一点，《科学》杂志最近发表的一篇论文表明，儿童在出生后一年就能够进行逻辑推理。请记住，先天并不意味着出生时就有。我长胡子的能力不是出生时就有的，而是与荷尔蒙和青春期有关。人类大脑的很大一部分实际上是在子宫外发育的，但发育时间相对较早。

GARY MARCUS: That’s right. My research with eight-month-olds also bears that out, and a recent paper in Science suggests that children are able to do logical reasoning after just the first year of life. Keep in mind that innate doesn’t mean exactly at birth. My ability to grow a beard is not something I began at birth, it was timed to hormones and puberty. A lot of the human brain actually develops outside the womb, but relatively early in life.

如果你观察马等早熟物种，它们几乎一出生就能走路，而且它们有相当复杂的视觉和障碍物探测能力。人类的一些机制在生命的第一年就形成了。你经常会听到人们说婴儿学会走路，但我认为事实并非如此。当然，有一些学习和肌肉力量校准等，但其中一些是成熟。一个包含完全发育的人类大脑的头部太大了，无法通过产道。

If you look at precocial species like horses, they can walk almost right away after being born, and they have fairly sophisticated vision and obstacle detection. Some of those mechanisms for humans get wired up in the first year of life. You’ll often hear people say that a baby learns to walk, but I don’t think that’s actually the case. There’s certainly some learning and calibration of muscle forces and so on, but some of it is maturation. A head containing a fully developed human brain would be too big to pass through the birth canal.

马丁·福特：即使你天生就有行走的能力，你也必须等待肌肉发育成熟后才能运用它。

MARTIN FORD: Even if you had an innate ability to walk, you’d have to wait for the muscles to develop before you could put it into operation.

加里·马库斯：是的，而且这些也没有完全发育。我们出生时还没有完全孵化，我认为这让人们感到困惑。头几个月发生的很多事情仍然在很大程度上受基因控制。这与学习本身无关。

GARY MARCUS: Right, and those aren’t fully developed either. We come out not quite fully hatched, and I think that confuses people. A lot of what’s going on in the first few months is still pretty much genetically controlled. It’s not about learning per se.

看看一只小野山羊。几天后，它就能爬下山了。它不是通过反复试验来学习这一点的——如果它从山上掉下来，它就会死——但它却能做出令人惊叹的导航和运动控制壮举。

Look at a baby ibex. After a couple of days, it can scramble down the side of a mountain. It doesn’t learn that by trial-and-error—if it falls off the side of a mountain then it’s dead—yet it can do spectacular feats of navigation and motor control.

我认为我们的基因组为我们大脑的运作方式勾画出了一个非常丰富的初稿，然后在此基础上进行了大量学习。当然，初稿的一部分内容是建立学习机制本身。

I think our genomes wire a very rich first draft of how our brains should operate, then there’s lots of learning on top of that. Some of that first draft is, of course, about making the learning mechanisms themselves.

人工智能领域的人们经常试图用尽可能少的先验知识来构建事物，我认为这是愚蠢的。事实上，科学家和普通人已经收集了大量关于世界的知识，我们应该将这些知识融入我们的人工智能系统中，而不是毫无理由地坚持从头开始。

People in AI often try to build things with as little prior knowledge as they can get away with, and I think that’s foolish. There’s actually lots of knowledge about the world that’s been gathered by scientists and ordinary people that we should be building into our AI systems, instead of insisting, for no really good reason, that we should start from scratch.

马丁·福特：大脑中存在的任何天赋必然是进化的结果，因此，对于人工智能来说，你可以对该天赋进行硬编码，或者你可以使用进化算法来自动生成它。

MARTIN FORD: Any innateness that exists in the brain has to be the result of evolution, so with an AI you could either hardcode that innateness, or perhaps you could use an evolutionary algorithm to generate it automatically.

加里·马库斯：这个想法的问题在于，进化过程非常缓慢且效率低下。进化需要数万亿生物和数十亿年的时间才能取得很好的结果。目前还不清楚在合理的时间内，实验室中的进化过程是否能取得足够的进展。

GARY MARCUS: The problem with that idea is that evolution is pretty slow and inefficient. It works over trillions of organisms and billions of years to get great results. It’s not clear that you’d get far enough with evolution in a lab in a reasonable timeframe.

思考这个问题的一种方式是，最初的 9 亿年进化过程并不那么令人兴奋。大多数细菌都是不同版本的，这并不那么令人兴奋。无意冒犯细菌。

One way to think about this problem is that the first 900 million years of evolution were not that exciting. Mostly you had different versions of bacteria, which is not that exciting. No offense to the bacteria.

然后，事情突然加速发展，出现了脊椎动物，然后是哺乳动物，然后是灵长类动物，最后出现了我们人类。进化速度加快的原因是，这就像在编程中拥有更多的子程序和更多的库代码。子程序越多，你就能越快地在此基础上构建更复杂的东西。在灵长类动物的大脑上构建一个有 100 或 1,000 个重要基因变化的人类是一回事，但你无法从细菌到人类大脑实现类似的飞跃。

Then suddenly things pick up and you get vertebrates, then mammals, then primates, and finally, you get us. The reason that the pace of evolution increased is because it’s like having more subroutines and more library code in your programming. The more subroutines you have, the quicker you can build more complicated things on top of that. It’s one thing to build a human on top of a primate brain with 100 or 1,000 important genetic changes, but you wouldn’t be able to make a similar leap from bacteria to a human brain.

研究进化神经网络的人往往从最基本的东西入手。他们试图进化单个神经元及其之间的连接，而我认为，在生物进化中，比如人类，已经有了非常复杂的遗传程序集。从本质上讲，你有一系列基因可以操作，但人们还没有真正弄清楚如何在进化编程环境中做到这一点。

People working on evolutionary neural networks often start too close to the bone. They’re trying to evolve individual neurons and connections between them, when my belief is that in the biological evolution of, say, humans, you already had very sophisticated sets of genetic routines. Essentially, you’ve got cascades of genes on which to operate and people haven’t really figured out how to do that in the evolutionary programming context.

我认为他们最终会实现，但部分原因是出于偏见，他们至今还没有实现。偏见是，“我想在实验室里从头开始，通过在七天内创造这个来证明我可以成为上帝。”这太荒谬了；这不会发生。

I think they will eventually, but partly because of prejudice they haven’t so far. The prejudice is, “I want to start from scratch in my lab and show that I can be God by creating this in seven days.” That’s ridiculous; it’s not going to happen.

马丁·福特：如果您要将这种先天性融入人工智能系统，您是否知道它会是什么样子？

MARTIN FORD: If you were going to build this innateness into an AI system, do you have a sense of what that would look like?

加里·马库斯：这包括两个部分。一个是功能上应该做什么，另一个是机械上应该如何做。

GARY MARCUS: There are two parts to it. One is functionally what it should do, and the other is mechanically how you should do it.

在功能层面，我有一些明确的建议，这些建议来自我自己的工作以及哈佛大学的伊丽莎白·斯佩尔克 (Elizabeth Spelke) 的工作。我在 2018 年初写的一篇论文中阐述了这一点，其中我谈到了十种不同的要求 ( https://arxiv.org/abs/1801.05667 )。我不会在这里深入讨论它们，但诸如符号操作和表示抽象变量的能力（计算机程序就是基于这些抽象变量的）；对这些变量的操作（计算机程序就是这些变量的）；类型标记区别，识别这个瓶子而不是一般的瓶子；因果关系；空间平移或平移不变性；物体倾向于沿着空间和时间上相连的路径移动的知识；认识到存在事物、地点等的集合。

At the functional level, I have some clear proposals drawing from my own work, and that of Elizabeth Spelke at Harvard. I laid this out in a paper that I wrote early in 2018, where I talked about ten different things that would be required (https://arxiv.org/abs/1801.05667). I won’t go into them in depth here, but things like symbol manipulation and the ability to represent abstract variables, which computer programs are based on; operations over those variables, which is what computer programs are; a type-token distinction, recognizing this bottle as opposed to bottles in general; causality; spatial translation or translation invariance; the knowledge that objects tend to move on paths that are connected in space and time; the realization that there are sets of things, places, and so on.

如果你有这样的东西，那么你就可以了解特定类型的物体在特定的地方以及被特定类型的代理操纵时会做什么。这比仅仅从像素中学习一切要好，后者是一种非常流行但我认为最终不够好的想法，这是我们目前在该领域看到的想法。

If you had things like that, then you could learn about what particular kinds of objects do when they’re in particular kinds of places and they’re manipulated by particular kinds of agents. That would be better than just learning everything from pixels, which is a very popular but I think ultimately inadequate idea that we are seeing in the field right now.

我们目前看到的是人们对诸如 Atari 游戏Breakout之类的像素进行深度强化学习，虽然你得到的结果看起来令人印象深刻，但它们却极其脆弱。

What we see at the moment is people doing deep reinforcement learning over pixels of, for example, the Atari game Breakout, and while you get results that look impressive, they’re incredibly fragile.

DeepMind 训练了一个 AI 来玩 Breakout，当你观看它时，你会发现它表现得非常棒。它应该学会了如何突破墙壁并将球困在顶部，这样它就可以反弹穿过许多方块。但是，如果你将挡板向上移动三个像素，整个系统就会崩溃，因为它实际上并不知道什么是墙，什么是反弹。它实际上只是学习了偶然事件，并在它记忆的偶然事件之间进行插值。程序并没有学习你所需要的抽象概念，而这正是从像素和非常低级的表示中完成所有事情的问题所在。

DeepMind trained an AI to play Breakout, and when you watch it, it looks like it’s doing great. It’s supposedly learned the concept of breaking through the wall and trapping the ball at the top so it can ricochet across a lot of blocks. However, if you were to move the paddle three pixels up, the whole system breaks because it doesn’t really know what a wall is or what a ricochet is. It’s really just learned contingencies, and it’s interpolating between the contingencies that it’s memorized. The programs are not learning the abstraction that you need, and this is the problem with doing everything from pixels and very low-level representations.

马丁·福特：理解物体和概念需要更高层次的抽象。

MARTIN FORD: It needs a higher level of abstraction to understand objects and concepts.

加里·马库斯：没错。你可能还需要真正地建立某些概念，比如“物体”。一种思考方式是将其视为学习处理颜色的能力。你不会从黑白视觉开始，最终了解到有颜色。它始于拥有两种不同的颜色受体色素，它们对光谱的特定部分敏感。然后，从那里，你可以了解特定的颜色。你需要先掌握一些先天的知识，然后才能完成其余的事情。也许以类似的方式，你可能需要天生就有一个物体的概念，也许还要约束物体不会随机出现和消失。

GARY MARCUS: Exactly. You may also need to actually build in certain notions. like “object.” One way to think about it is like the ability to learn to process color. You don’t start with black-and-white vision and eventually learn that there is color. It starts by having two different color-receptor pigments that are sensitive to particular parts of the spectrum. Then, from there, you can learn about particular colors. You need some piece to be innate before you can do the rest. Maybe in a similar way, you might need to have innately the notion that there’s an object, and maybe the constraint that objects don’t just randomly appear and disappear.

想象一下这样一个世界，那里有一台星际迷航传送器，任何东西都可能在任何时刻出现在任何地方。你永远无法从中学到东西。让我们了解世界的原因在于，物体确实在空间和时间上相连的路径上移动，而十亿年的进化可能是为了让你更快地离开地面而建立的。

Imagine a world in which there was a Star Trek transporter, and anything could appear at any place at any moment. You’d never be able to learn from that. What allows us to learn about the world is the fact that objects do move on paths that are connected in space and time, and over a billion years of evolution that might have been wired in as a way of getting you off the ground faster.

马丁·福特：我们来谈谈未来吧。您认为实现 AGI 的主要障碍是什么？我们能用现有的工具实现吗？

MARTIN FORD: Let’s talk about the future. What do you see as the main hurdles to getting to AGI, and can we get there with current tools?

GARY MARCUS：我认为深度学习是进行模式分类的有用工具，这是任何智能代理都需要解决的一个问题。我们要么保留它，要么用能更有效地完成类似工作的东西来代替它，我认为这是可能的。

GARY MARCUS: I see deep learning as a useful tool for doing pattern classification, which is one problem that any intelligent agent needs to do. We should either keep it around for that, or replace it with something that does similar work more efficiently, which I do think is possible.

与此同时，智能代理还需要做其他一些事情，而深度学习目前并不擅长这些事情。它不擅长抽象推理，也不是一个很好的语言工具，除非是翻译之类的不需要真正理解的事情，或者至少不需要做近似翻译。它也不擅长处理以前没有见过的情况，也不擅长处理信息相对不完整的情况。因此，我们需要用其他工具来补充深度学习。

At the same time, there are other kinds of things that intelligent agents need to do that deep learning is not currently very good at. It’s not very good at abstract inference, and it’s not a very good tool for language, except things like translation where you don’t need real comprehension, or at least not to do approximate translation. It’s also not very good at handling situations that it hasn’t seen before and where it has relatively incomplete information. We therefore need to supplement deep learning with other tools.

更广泛地说，人类对世界的很多知识都可以通过符号来编码，无论是通过数学还是语言中的句子。我们确实希望将这些符号信息与其他更具感知性的信息结合起来。

More generally, there’s a lot of knowledge that humans have about the world that can be codified symbolically, either through math or sentences in a language. We really want to bring that symbolic information together with the other information that’s more perceptual.

心理学家讨论了自上而下的信息和自下而上的信息之间的关系。如果你看一张图片，光线会落在你的视网膜上，这就是自下而上的信息，但你也会利用你对世界的了解和对事物行为的经验，将自上而下的信息添加到你对图片的解读中。

Psychologists talk about the relationship between top-down information and bottom-up information. If you look at an image, light falls on your retina and that’s bottom-up information, but you also use your knowledge of the world and your experience of how things behave to add top-down information to your interpretation of the image.

深度学习系统目前专注于自下而上的信息。它们可以解释图像的像素，但对图像中包含的对象一无所知。

Deep learning systems currently focus on bottom-up information. They can interpret the pixels of an image, but don’t then have any knowledge of the object the image contains.

最近的一个例子是Adversarial Patch 论文（https://arxiv.org/pdf/1712.09665.pdf）。在论文中，他们展示了如何通过在图像上添加贴纸来欺骗深度学习系统。他们拍摄了一张深度学习系统非常有信心识别的香蕉照片，然后在照片中的香蕉旁边添加了一个看起来像迷幻烤面包机的贴纸。任何人看到它都会说这是一根香蕉，旁边有一个看起来很有趣的贴纸，但深度学习系统会立即非常有信心地说，它现在是一张烤面包机的照片。

An example of this recently was the Adversarial Patch paper (https://arxiv.org/pdf/1712.09665.pdf). In the paper, they show how you can fool a deep learning system by adding a sticker to an image. They take a photo of a banana that is recognized with great confidence by a deep learning system and then add a sticker that looks like a psychedelic toaster next to the banana in the photo. Any human looking at it would say it was a banana with a funny looking sticker next to it, but the deep learning system immediately says, with great confidence, that it’s now a picture of a toaster.

深度学习系统只是试图说出图像中最突出的东西是什么，高对比度的迷幻烤面包机吸引了它的注意力，而它忽略了非常清晰的香蕉。

The deep learning system is just trying to say what the most salient thing in the image is, and the high-contrast psychedelic toaster grabs its attention and it ignores the perfectly clear banana.

这是深度学习系统只获取自下而上信息的例子，也就是枕叶皮质的功能。它根本无法捕捉额叶皮质在推理真正发生的事情时所做的事情。

This is an example of how deep learning systems are only getting the bottom-up information, which is what your occipital cortex does. It’s not capturing at all what your frontal cortex does when it reasons about what’s really going on.

为了实现 AGI，我们需要能够捕捉到等式的两边。换句话说，人类拥有各种常识推理能力，这必须是解决方案的一部分。深度学习并没有很好地捕捉到这一点。在我看来，我们需要将人工智能中历史悠久的符号操作与深度学习结合起来。它们被分开处理的时间太长了，现在是时候将它们结合起来了。

To get to AGI, we need to be able to capture both sides of that equation. Another way to put it is that humans have all kinds of common-sense reasoning, and that has to be part of the solution. It’s not well captured by deep learning. In my view, we need to bring together symbol manipulation, which has a strong history in AI, with deep learning. They have been treated separately for too long, and it’s time to bring them together.

马丁·福特：如果您必须指出目前正在进行的、最接近 AGI 的一家公司或项目，您会指出谁？

MARTIN FORD: If you had to point to one company or project that’s going on now that is the closest to being on the path to AGI, who would you point to?

加里·马库斯：我对艾伦人工智能研究所的马赛克项目感到非常兴奋。他们正在再次尝试解决道格·莱纳特试图解决的问题，即如何将人类知识转化为可计算的形式。这并不是要回答诸如巴拉克·奥巴马出生在哪里之类的问题——计算机实际上可以很好地表示这些信息，并可以从可用数据中提取这些信息。

GARY MARCUS: I’m very excited about Project Mosaic at the Allen Institute for AI. They’re taking a second crack at the problem Doug Lenat was trying to solve, which was how you take human knowledge and put it in computable form. This is not about answering questions like where was Barack Obama born—computers actually represent that information pretty well and can extract it from available data.

不过，还有很多信息没有写在任何地方，例如，烤面包机比汽车小。维基百科上很可能没有人这么说，但我们知道这是真的，这让我们能够做出推断。如果我说“加里被烤面包机碾过了”，你会觉得这很奇怪，因为烤面包机并不是那么大的物体，但“被汽车碾过了”是有道理的。

There’s a lot of information, though, that’s not written anywhere, for example, toasters are smaller than cars. The likelihood is that no one says that on Wikipedia, but we know it to be true and that allows us to make inferences. If I said, “Gary got run over by a toaster,” you would think that’s weird because a toaster’s not that big an object, but “run over by a car” makes sense.

马丁·福特：那么这是属于符号逻辑领域吗？

MARTIN FORD: So this is in the area of symbolic logic?

加里·马库斯：嗯，有两个相关的问题。一个问题是，你如何获得这些知识？另一个问题是，你想用符号逻辑来操纵这些知识吗？

GARY MARCUS: Well, there are two related questions. One question is, how do you get that knowledge at all? The other is, do you want symbolic logic as a way to manipulate that?

我最好的猜测是符号逻辑实际上对此非常有用，我们不应该将其抛弃。我愿意接受有人找到另一种方法来处理它，但我没有看到人们已经很好地处理了它，而且我看不出我们如何在不具备一些常识的情况下构建真正理解语言的系统。这是因为每次我对你说一句话时，你都会有一些常识性的知识来理解这句话。

My best guess is that symbolic logic is actually pretty useful for it, and we shouldn’t throw it out the window. I’m open to somebody finding another way to deal with it, but I don’t see any ways in which people have dealt with it well, and I don’t see how we can build systems that really understand language without having some of that common sense. This is because every time I say a sentence to you, there’s some common-sense knowledge that goes into your understanding of that sentence.

如果我告诉你我要骑自行车从纽约到波士顿，我不必告诉你我不会飞过天空、潜入水下或绕道去加利福尼亚。你可以自己弄清楚所有这些事情。这不是字面意思，而是你对人类的了解，他们喜欢走高效的路线。

If I tell you I’m going to ride my bicycle from New York to Boston, I don’t have to tell you that I’m not going to fly through the air, go underwater, or take a detour to California. You can figure all that stuff out for yourself. It’s not in the literal sentence, but it’s in your knowledge about humans that they like to take efficient routes.

作为人类，你可以做出很多推论。如果不填写这些推论，你就无法理解我的句子，而你实际上是在阅读字里行间的内容。我们读懂了大量字里行间的内容，但要使整个交易顺利进行，必须有共同的常识，而我们还没有拥有这种共同常识的机器。

You, as a human, can make a lot of inferences. There’s no way you can understand my sentences without filling in those inferences, where you’re effectively reading between the lines. We read an enormous amount between the lines, but for that whole transaction to work there has to be shared common sense, and we don’t have machines that have that shared common sense yet.

最大的项目是道格·莱纳特（Doug Lenat）的 Cyc，该项目始于 1984 年左右，据大多数人说，它的效果并不好。它是 30 年前以封闭形式开发的。如今，我们对机器学习有了更多的了解，艾伦人工智能研究所致力于以开源方式开展工作，让社区可以参与其中。我们现在对大数据的了解比 20 世纪 80 年代更多，但这仍然是一个非常困难的问题。重要的是，当其他人都在回避它时，他们正在面对它。

The biggest project to do that was Doug Lenat’s Cyc, which started around 1984 and by most accounts, it didn’t work very effectively. It was developed 30 years ago in a closed form. Nowadays we know much more about machine learning, and the Allen Institute for AI is committed to doing things in open-source ways in which the community can participate. We know more about big data now than we did back in the 1980s, but it’s still a very difficult problem. The important thing is that they’re confronting it when everybody else is hiding from it.

马丁·福特：您认为 AGI 的时间表是怎样的？

MARTIN FORD: What do you think the timeframe is for AGI?

加里·马库斯：我不知道。我知道目前还没有实现这一目标的大部分原因以及需要解决的问题，但我认为你无法确定一个具体日期。我认为你需要一个置信区间——统计学家会这样描述——围绕它。

GARY MARCUS: I don’t know. I know most of the reasons why it’s not here now and the things that need to be solved, but I don’t think you can put a single date on that. What I think is that you need a confidence interval—as a statistician would describe it—around it.

我可能会告诉你，如果我们非常幸运的话，我认为它会在 2030 年左右到来，更可能是 2050 年，或者在最坏的情况下是 2130 年。关键是很难给出一个确切的日期。有很多事情我们不知道。我总是想起比尔盖茨在 1994 年写的《未来之路》一书，甚至他也没有真正意识到互联网会改变一切。我的观点是，可能会有各种我们没有预料到的事情发生。

I might tell you that I think it’ll come between 2030 if we’re phenomenally lucky and more likely 2050, or in the worst case 2130. The point is that it’s very hard to give an exact date. There are lots of things we just don’t know. I always think about how Bill Gates wrote the book The Road Ahead in 1994, and even he didn’t really realize that the internet was going to change things as it did. My point is that there could be all kinds of things that we’re just not anticipating.

目前，机器的智能还很弱，但我们不知道人类下一步会发明什么。大量资金投入该领域，可能会推动事情的发展，或者也可能比我们想象的要困难得多。我们真的不知道。

Right now, machines are weak at intelligence, but we don’t know what people are going to invent next. There’s a lot of money going into the field, which could move things along, or alternatively it could be much harder than we think that it is. We just really don’t know.

马丁·福特：这仍然是一个相当激进的时间框架。您建议最早在 12 年后，或者最远在 112 年后。

MARTIN FORD: That’s still a fairly aggressive time frame. You’re suggesting as soon as 12 years away or as far away as 112 years.

加里·马库斯：当然，这些数字也可能是错误的。从另一个角度来看，虽然我们在狭义智能方面取得了很大进展，但迄今为止我们在通用智能（AGI）方面还没有取得那么大的进展。

GARY MARCUS: And those figures could of course be wrong. Another way to look at it is while we’ve made a lot of progress on narrow intelligence, we haven’t made nearly as much progress so far on general intelligence, AGI.

苹果的 Siri 于 2010 年问世，其工作原理与 1964 年创建的早期自然语言计算机程序 ELIZA 并无太大区别，后者通过匹配模板来给人一种它理解语言的错觉，但实际上它并不能理解。我的乐观情绪很大程度上来自于有多少人在努力解决这个问题，以及有多少企业投入了多少资金来尝试解决这个问题。

Apple’s Siri, which began life in 2010, doesn’t work that much differently from ELIZA, an early natural language computer program created in 1964, which matched templates in order to give an illusion of understanding language that it didn’t really. A lot of my optimism comes from how many people are working at the problem and how much money businesses are investing to try and solve it.

马丁·福特：这绝对是一个巨大的变化，因为人工智能不再只是大学里的一个研究项目。现在人工智能已经成为谷歌和 Facebook 等大公司商业模式的核心。

MARTIN FORD: It definitely is a massive change in terms of AI not being something just done as a research project at a university. Now AI is central to the business models of big companies like Google and Facebook.

加里·马库斯：人工智能领域投入的资金远远超过以往任何时候，尽管在 20 世纪 60 年代和 70 年代初期，也就是所谓的“人工智能寒冬”到来之前，投入的资金确实很多。同样重要的是，要认识到，资金并不是解决人工智能问题的保证，但很可能是先决条件。

GARY MARCUS: The amount of money being spent on AI far eclipses anything before, although there certainly was a lot of money spent in the 1960s and early 1970s before the first so-called AI Winter. It’s also important to acknowledge that money is not a guaranteed solution to the problems of AI, but is very likely a prerequisite.

马丁·福特：让我们关注一项更为狭窄的技术的预测：自动驾驶汽车。

MARTIN FORD: Let’s focus on a prediction for a much narrower technology: the self-driving car.

我们什么时候才能呼叫像 Uber 这样的汽车，它完全由人工智能驾驶，可以在随机地点接你，然后带你到你指定的目的地？

When are we going to be able to call for something like an Uber that’s driven by nothing but an AI, and that can pick you up at a random location and then take you to a destination that you specify?

加里·马库斯：这至少还需要十年，甚至可能更久。

GARY MARCUS: It’s at least a decade away, and probably more.

马丁·福特：你几乎进入了与你的 AGI 预测相同的领域。

Martin Ford: You’re almost getting into the same territory as your AGI prediction.

加里·马库斯：没错，主要原因是，如果你谈论的是像曼哈顿或孟买这样交通繁忙的大都市，那么人工智能将面临很多不可预测性。在凤凰城驾驶无人驾驶汽车是一回事，那里的天气很好，人口密度也低得多。曼哈顿的问题是，任何事情都可能发生，没有人特别守规矩，每个人都很激进，发生不可预测事件的可能性要高得多。

GARY MARCUS: That’s right, and for the principal reason that if you’re talking about driving in a very heavy metropolitan location like Manhattan or Mumbai, then the AI will face a lot of unpredictability. It’s one thing to have a driverless car in Phoenix, where the weather is good and the population is a lot less densely packed. The problem in Manhattan is that anything goes at any moment, nobody is particularly well-behaved and everybody is aggressive, the chance of having unpredictable things occur is much higher.

即使是简单的道路元素，比如保护行人的路障，也会给人工智能带来问题。这些都是人类通过推理处理的复杂情况。目前，无人驾驶汽车依靠高度详细的地图和激光雷达等设备进行导航，但无法真正理解其他驾驶员的动机和行为。人类的视觉系统还行，但对外面的情况和自己开车时的行为有很好的理解。机器正试图利用大数据来绕过这一点，我不清楚，仅仅通过向这些大数据驱动的系统添加更多数据，是否就能达到在曼哈顿驾驶所需的准确度。你的准确度可能会达到 99.99%，但如果你用数字来衡量，那比人类差多了，而且在路上使用这种规模的机器太危险了，尤其是在曼哈顿这样的繁忙街道上。

Even simple road elements like barricades to protect people can cause issues for an AI. These are complex situations that humans deal with by using reasoning. Right now, driverless cars navigate by having highly detailed maps and things like LIDAR, but no real understanding of the motives and behavior of other drivers. Humans have an OK visual system, but a good understanding of what’s out there and what they’re doing when they’re driving. Machines are trying to fake their way around that with big data, and it’s not clear to me that you get to the accuracy levels you need for driving in Manhattan simply by adding more data to these big data-driven systems. You might get to 99.99% accuracy, but if you do the numbers on that, that’s much worse than humans, and it’s far too dangerous to have that scale on the road, especially on busy streets like in Manhattan.

马丁·福特：那么也许有一个更短期的解决方案，它不是带你去你选择的地点，而是带你去一个预先定义好的位置？

MARTIN FORD: Maybe then there’s a nearer term solution, where rather than a location of your choice, it takes you to a predefined location?

加里·马库斯：我们可能很快就能在凤凰城或其他有限的地点实现这一点。如果你能找到一条永远不需要左转的路线，那里不太可能有人挡道，而且交通也很文明，你也许可以做到这一点。我们已经在机场安装了单轨列车，它们按照预先定义的路径以类似的方式运行。

GARY MARCUS: There’s a possibility that very soon we may well have that in Phoenix, or another limited location. If you can find a route where you never need to take a left turn, where humans are unlikely to be in the way, and the traffic is civilized, you might be able to do that. We already have monorails at airports that work in a similar way following a predefined path.

这里有一系列的危险，从机场单轨铁路等严密控制的环境（轨道上不允许任何人）到曼哈顿街道（任何时间都有可能出现任何人和任何事物）。我们还有其他因素，比如天气比凤凰城复杂得多。我们什么都有。雨夹雪、雪泥、冰雹、树叶、卡车上掉下来的东西；什么都有。

There’s a continuum from super-controlled circumstances like the monorail at the airport where nobody should be on that track, to the Manhattan streets where anybody and anything could be there at any time. We also have other factors like the fact that weather is much more complicated than in Phoenix. We get everything. Sleet, slush, hail, leaves, things that fall off trucks; everything.

你越深入一个无界的开放式系统，面临的挑战就越大，你就越需要能够像 AGI 系统那样进行推理。它仍然不像 AGI 本身那样开放，但它开始接近这一点，这就是为什么我的数字并没有完全不同。

The more you go into an unbounded open-ended system, the more challenge there is and the more you need to be able to reason the way an AGI system does. It’s still not as open-ended as AGI per se, but it starts to approach that, and that’s why my numbers are not totally different.

马丁·福特：让我们来谈谈我关注的一个领域：人工智能对经济和就业市场的影响。

MARTIN FORD: Let’s talk about an area that I’ve focused on a lot: the economic and job market impact of AI.

许多人认为我们正处于新工业革命的前沿；这场革命将彻底改变劳动力市场的面貌。您同意这个观点吗？

Many people believe we’re on the leading edge of a new Industrial Revolution; something that’s going to completely change the way the labor market looks. Would you agree with that?

加里·马库斯：我确实同意这一点，不过时间上会稍微慢一些。无人驾驶汽车比我们想象的要难，所以付费司机暂时是安全的，但快餐店员工和收银员却深陷困境，工作场所中这样的人很多。我确实认为这些根本性的变化将会发生。其中一些变化会很慢，但从 100 年的规模来看，如果某件事需要额外的 20 年，那就不算什么了。

GARY MARCUS: I do agree with that, though on a slightly slower time frame. Driverless cars are harder than we thought, so paid drivers are safe for a while, but fast-food workers and cashiers are in deep trouble, and there’s a lot of them in the workplace. I do think these fundamental changes are going to happen. Some of them will be slow, but in the scale of, say, 100 years, if something takes an extra 20 years, it’s nothing.

无论是 2030 年还是 2070 年，本世纪的某个时候，人工智能机器人和就业将会出现问题。在某个时候，我们需要改变我们的社会结构，因为我们将会到达一个就业岗位减少但劳动年龄人口仍然充足的阶段。

There is going to be a problem with AI robots and employment sometime in this century, whether it’s 2030 or 2070. At some point we need to change how we structure our societies because we are going to get to a point where there’s less employment available but still a working-age population.

也有反对意见，比如大多数农业工作消失后就被工业工作取代了，但我觉得这些理由并不令人信服。我们面临的主要问题是规模，以及一旦你有了解决方案，就可以在任何地方以相对低廉的成本使用它。

There are counterarguments, like when most agricultural jobs disappeared they were just replaced by industrial jobs, but I don’t find them to be compelling. The main issue we will face is scale and the way that once you have a solution, you can use it everywhere relatively cheaply.

开发第一个有效的无人驾驶汽车算法/数据库系统可能需要 50 年的时间，并耗费数十亿美元的研究资金，但一旦我们拥有它们，人们就会大规模推广它们。一旦我们达到这一点，数百万卡车司机将面临在几年内失去工作的境地。

Getting the first driverless car algorithm/database system that works might be 50 years of work and cost billions of dollars in research, but once we have them, people are going to roll them out at scale. As soon as we reach that point, millions of truck drivers are going to be put in a position where they could lose their jobs within a matter of years.

目前还不清楚我们是否会出现新的工作岗位，以取代卡车司机行业的现有工作岗位。许多新出现的工作岗位需要更少的人手。例如，YouTube 创业者是一个很棒的工作。你可以在家里制作视频赚取数百万美元。这很棒，但也许只有 1,000 人这样做，而不是 100 万人，而且不足以取代所有可能消失的卡车司机工作岗位。

It’s not clear that we are going to have new jobs appear that can replace existing jobs at the scale of the truck-driving industry. A lot of the new jobs that have arisen need fewer people. For example, a YouTube entrepreneur is a great job. You can make millions of dollars staying at home making videos. That’s terrific, but maybe 1,000 people do that, not a million people and not enough to replace all the potentially lost truck driving jobs.

我们很容易想象到未来会出现一些以前没有过的工作，但在这个仅靠 18 个人就能创建 Instagram 之类的网站的时代，要出现能够雇用大量人员的新兴行业却很难。

It’s easy to come up with jobs that we will have that we didn’t have before, but it’s hard to come up with new industries that will employ large numbers of people in an era where you can build something like Instagram with 18 people.

马丁·福特：目前大约有一半的劳动力从事的是基本可预测的活动。他们所做的事情都包含在数据中，最终会在一段时间内受到机器学习的影响。

MARTIN FORD: Probably something like half the current workforce is engaged in fundamentally predictable activities. What they’re doing is encapsulated in the data and ultimately is going to be susceptible to machine learning over some time frame.

加里·马库斯：没错，但目前这类事情相当困难。人工智能系统无法像人类一样真正理解自然语言数据。例如，从医疗记录中提取信息是机器目前很难做到的事情。这是可以预测的，也不是那么难，但用机器做好这件事还需要一段时间。不过从长远来看，我同意你的观点。自然语言理解会越来越好，最终那些可预测的工作将会消失。

GARY MARCUS: True, but things like that that are pretty hard right now. AI systems just don’t really understand the data as a natural language like humans do. For example, extricating information from medical records is something that’s very hard for machines to do right now. It’s predictable, and it’s not that hard work, but doing it well with a machine is a while away. In the long run, though, I agree with you. Natural language understanding will get better and eventually those predictable jobs will go away.

马丁·福特：鉴于您认为人工智能终有一天会造成失业，您是否支持将基本收入作为解决这一问题的潜在解决方案？

MARTIN FORD: Given that you believe job loss to AI will happen at some point, would you support a basic income as a potential solution to that?

加里·马库斯：我认为没有其他选择。我们会实现目标，但问题是我们是通过一项普遍协议和平实现目标，还是会发生街头骚乱和人员死亡。我不知道方法是什么，但我看不到其他结局。

GARY MARCUS: I see no real alternative. We will get there, but it’s a question of whether we get there peacefully through a universal agreement or whether there are riots on the street and people getting killed. I don’t know the method, but I don’t see any other ending.

马丁·福特：你可以说技术已经产生了这种影响。目前美国确实存在阿片类药物泛滥，工厂自动化技术可能在其中发挥了一定作用，导致中产阶级就业机会消失。也许阿片类药物的使用与某些人（尤其是工薪阶层男性）感到的尊严丧失甚至绝望有关？

MARTIN FORD: You could argue that technology is already having an impact of that sort. We do have an opioid epidemic in the US at the moment, and automation technology in factories has likely played a role in that in terms of middle-class job opportunities disappearing. Perhaps opioid use is tied to a perceived loss of dignity or even despair among some people, especially working-class men?

加里·马库斯：我会谨慎地做出这个假设。这可能是真的，但我不认为其中的联系是牢不可破的。在我看来，一个更好的比喻是，很多人把手机当作阿片类药物，智能手机是人们的新鸦片。我们可能正在走向一个很多人只是沉迷于虚拟现实的世界，如果经济状况良好，他们可能会相当高兴。我不确定这一切会走向何方。

GARY MARCUS: I would be careful about making that assumption. It may be true, but I don’t think the links are ironclad. A better analogy, in my opinion, is how a lot of people use their phones as an opioid, and that smartphones are the new opium of the people. We may be moving toward a world where a lot of people just hang out in virtual reality, and if the economics work they may be reasonably happy. I’m not sure where that’s all going to go.

马丁·福特：人们提出了一系列与人工智能相关的风险。伊隆·马斯克等人尤其直言不讳地指出了生存威胁。您认为我们应该担心人工智能的影响和风险吗？

MARTIN FORD: There’s a range of risks associated with AI that have been raised. People like Elon Musk have been especially vocal about existential threats. What do you think we should be worried about in terms of the impacts and risks of AI?

加里·马库斯：我们应该担心人们会以恶意的方式使用人工智能。真正的问题是，随着人工智能越来越深入电网，越来越容易受到黑客攻击，人们会如何利用人工智能所拥有的力量。我并不担心人工智能系统会独立地想把我们当早餐吃掉，或者把我们变成回形针。这并非完全不可能，但没有真正的证据表明我们正朝着这个方向发展。不过，有证据表明，我们正赋予这些机器越来越多的权力，而我们不知道如何在短期内解决网络安全威胁。

GARY MARCUS: We should be worrying about people using AI in malevolent ways. The real problem is what people might do with the power that AI holds as it becomes more embedded in the grid and more hackable. I’m not that worried about AI systems independently wanting to eat us for breakfast or turn us into paper clips. It’s not completely impossible, but there’s no real evidence that we’re moving in that direction. There is evidence, though, that we’re giving more and more power to those machines, and that we have no idea how to solve the cybersecurity threats in the near term.

马丁·福特：那么长期威胁呢？埃隆·马斯克和尼克·博斯特罗姆非常担心人工智能的控制问题；他们认为可能存在一个递归的自我改进循环，从而导致智能爆炸。你不能完全忽视这一点，对吧？

MARTIN FORD: What about long-term threats, though? Elon Musk and Nick Bostrom are very concerned about the control problem with AI; the idea that there could be a recursive self-improvement cycle that could lead to an intelligence explosion. You can’t completely discount that, right?

加里·马库斯：我并不完全否认这一点，我并不是说可能性为零，但短期内发生的可能性相当低。最近有一段机器人打开门把手的视频流传开来，这就是它们目前的发展水平。

GARY MARCUS: I don’t completely discount it, I’m not going to say the probability is zero but the probability of it happening anytime soon is pretty low. There was recently a video circulated of robots opening doorknobs, and that’s about where they are in development.

我们根本没有能够稳健地驾驭世界的人工智能系统，也没有知道如何改进自己的机器人系统，除非以有限的方式改进，例如将其电机控制系统定制为特定功能。这不是当前的问题。我认为在这个领域投资一些资金并让人们思考这些问题是可以的。我的问题是，正如我们在 2016 年美国大选中看到的那样，还有更紧迫的问题，比如使用人工智能来生成和定位虚假新闻。这是当今的问题。

We don’t have AI systems that can robustly navigate our world at all, nor do we have robotic systems that know how to improve themselves, except in constrained ways like tailoring their motor control system to a particular function. This is not a current problem. I think it’s fine to invest some money in the field and have some people think about those problems. My issue is that, as we saw with the 2016 US election, there are more pressing problems like using AI to generate and target fake news. That’s a problem today.

马丁·福特：您之前说过 AGI 最早可在 2030 年实现。如果一个系统真正智能，甚至可能超级智能，我们是否需要确保它的目标与我们希望它做的事情一致？

MARTIN FORD: Earlier you said AGI is conceivable as early as 2030. If a system is genuinely intelligent, potentially superintelligent, do we need to make sure that its goals are aligned with what we want it to do?

加里·马库斯：是的，我想是的。如果我们这么快就实现这个目标，我会感到惊讶，但正是因为这个原因，我认为我们应该让一些人考虑这些问题。我只是不认为它们是我们目前最紧迫的问题。即使我们真的实现了 AGI 系统，谁又能保证 AGI 系统会对干涉人类事务有任何兴趣呢？

GARY MARCUS: Yes, I think so. I will be surprised if we get there that quickly, but it’s for that reason why I think that we should have some people thinking about these problems. I just don’t think that they’re our most pressing problems right now. Even when we do get to an AGI system, who’s to say that the AGI system is going to have any interest whatsoever in meddling in human affairs?

大约 60 年前，人工智能无法在跳棋中取胜，而去年，人工智能已经能够在难度大得多的围棋中取胜。你可以绘制游戏智商图，并制作一个量表，表明游戏智商在 60 年内从 0 上升到了 60。然后，你可以对机器恶意行为做类似的事情。机器恶意行为在这段时间内根本没有变化。没有相关性，机器恶意行为为零。过去没有，现在也没有。这并不意味着它是不可能的——我不想做归纳论证说因为它从未发生过，所以它永远不会发生——但没有任何迹象表明这一点。

We’ve gone from AI not being able to win at checkers about 60 years ago to being able to win at Go, which is a much harder game, in the last year. You could plot a game IQ, and make up a scale to say that game IQ has gone from 0 to 60 in 60 years. You could then do a similar thing for machine malevolence. Machine malevolence has not changed at all over that time. There’s no correlation, there is zero machine malevolence. There was none, and there is none. It doesn’t mean that it’s impossible—I don’t want to make the inductive argument that because it never happened, it never will—but there’s no indication of it.

马丁·福特：在我看来，这听起来像是一个阈值问题，在拥有通用人工智能之前，机器不可能产生恶意。

MARTIN FORD: It sounds to me like a threshold problem, though, you can’t have machine malevolence until you have AGI.

加里·马库斯：有可能。其中一些与动机系统有关，你可以尝试构建一个论点，说 AGI 是机器恶意的先决条件，但你不能说它是必要和充分条件。

GARY MARCUS: Possibly. Some of it has to do with motivational systems and you could try to construct an argument saying that AGI is a prerequisite for machine malevolence, but you couldn’t say that it’s a necessary and sufficient condition.

这是一个思维实验。我可以说出一个基因因素，它将使你实施暴力行为的几率增加 5 倍。如果你没有这个基因因素，你的暴力倾向就很低。机器会不会有这个基因因素？当然，基因因素是男性与女性的对比。

Here’s a thought experiment. I can name a single genetic factor that will increase your chance of committing violent acts by a factor of 5. If you don’t have it, your proclivity to violence is pretty low. Are machine’s going to have this genetic factor, or not? The genetic factor, of course, is male versus female.

马丁·福特：这就是将 AGI 设定为女性的理由吗？

MARTIN FORD: Is that an argument for making AGI female?

加里·马库斯：性别只是一种替代，并不是真正的问题，但它是让人工智能变得非暴力的一个理由。我们应该制定限制和法规，以减少人工智能使用暴力或自行想出对我们做什么的想法的可能性。这些都是艰难而重要的问题，但它们并不像埃隆的言论可能让很多人认为的那样简单。

GARY MARCUS: The gender is a proxy, it’s not the real issue, but it is an argument for making AI nonviolent. We should have restrictions and regulations to reduce the chance of AI being violent or of coming up with ideas of its own about what it wants to do with us. These are hard and important questions, but they’re much less straightforward than Elon’s quotes might lead a lot of people to think.

马丁·福特：不过，他对 OpenAI 的投资听起来并不是坏事。应该有人来做这项工作。让政府投入大量资源解决人工智能控制问题很难说得通，但让私人实体来做这件事似乎是件好事。

MARTIN FORD: What he’s doing in terms of making an investment in OpenAI doesn’t sound like a bad thing, though. Somebody ought to be doing that work. It would be hard to justify having the government invest massive resources in working on AI control issues, but having private entities doing that seems positive.

加里·马库斯：美国国防部确实在这些方面投入了一些资金，这是理所应当的，但你必须有一个风险投资组合。我更担心某些类型的生物恐怖主义，而不是这些特定的人工智能威胁，我更担心网络战，这是一个真正持续存在的问题。

GARY MARCUS: The US Department of Defense does spend some money on these things, as they should, but you have to have a risk portfolio. I’m more worried about certain kinds of bioterrorism than I am about these particular AI threats, and I’m more worried about cyber warfare, which is a real going concern.

这里有两个关键问题。一是，您认为 X 的概率大于 0 吗？答案显然是肯定的。二是，相对于您可能担心的其他风险，您会将其排在第几位？对此，我想说，这些都是不太可能发生的情形，还有其他更可能发生的情形。

There are two key questions here. One is, do you think that the probability of X is greater than 0? The answer is clearly yes. The other is, relative to the other risks that you might be concerned about, where would you rank this? To which I would say, these are somewhat unlikely scenarios, and there are other scenarios that are more likely.

马丁·福特：如果我们在某个时候成功打造出通用人工智能，你认为它会有意识吗？或者有可能出现一个没有内在体验的智能僵尸吗？

MARTIN FORD: If at some point we succeed in building an AGI, do you think it would be conscious, or is it possible to have an intelligent zombie with no inner experience?

加里·马库斯：我认为是后者。我不认为意识是先决条件。它可能是人类或其他生物的附带现象。还有另一个思想实验说，我们是否有某种行为和我一样但没有意识的东西？我认为答案是肯定的。我们不确定，因为我们没有任何独立的衡量意识是什么的标准，所以很难为这些论点提供依据。

GARY MARCUS: I think it’s the latter. I don’t think that consciousness is a prerequisite. It might be an epiphenomenon in humans or maybe some other biological creatures. There’s another thought experiment that says, could we have something that behaves just like me but isn’t conscious? I think the answer is yes. We don’t know for sure, because we don’t have any independent measure of what consciousness is, so it’s very hard to ground these arguments.

我们如何判断机器是否有意识？我怎么知道你有意识？

How would we tell if a machine was conscious? How do I know that you’re conscious?

马丁·福特：好吧，你可以假设我是，因为我们是同一个物种。

MARTIN FORD: Well, you can assume I am because we’re the same species.

加里·马库斯：我认为这是一个错误的假设。如果事实证明意识是随机分布在人口中的四分之一的人身上，那会怎样？如果这只是一个基因，那会怎样？我有超级味觉基因，这使我对苦味化合物很敏感，但我的妻子没有。她看起来和我来自同一个物种，但我们在这种特性上有所不同，所以也许我们在意识特性上也有所不同？我在开玩笑，但在这里我们真的不能使用客观的衡量标准。

GARY MARCUS: I think that’s a bad assumption. What if it turns out that consciousness is randomly distributed through our population to one-quarter of the people? What if it’s just a gene? I have the supertaster gene that makes me sensitive to bitter compounds, but my wife doesn’t. She looks like she’s from the same species as me, but we differ in that property, and so maybe we differ in the consciousness property also? I’m kidding, but we can’t really use an objective measure, here.

马丁·福特：这听起来像是一个不可知的问题。

MARTIN FORD: It sounds like an unknowable problem.

加里·马库斯：也许有人会想出更聪明的答案，但到目前为止，大多数学术研究都集中在我们称为意识的意识部分。你的中枢神经系统在什么时候逻辑地意识到某些信息是可用的？

GARY MARCUS: Maybe someone will come up with a cleverer answer, but so far, most of the academic research is focused on the part of consciousness we call awareness. At what point does your central neural system realize logically that certain information is available?

研究表明，如果你只看到某个东西 100 毫秒，那么你可能没有意识到自己看到了它。如果你只看到它半秒钟，那么你很确定自己确实看到了它。有了这些数据，我们可以开始建立一种特征，即哪个神经回路在哪个时间范围内提供你可以反思的信息，我们可以称之为意识。我们正在取得进展，但还不是普遍意识。

Research has shown that if you only see something for 100 milliseconds then you might not realize you’d seen it. If you see it for half a second, you’re pretty sure you actually saw it. With that data we can start to build up a characterization of which neural circuits at which time frame contribute information that you can reflect on, and we can call that awareness. That we’re making progress on, but not yet general consciousness.

马丁·福特：你显然认为 AGI 是可以实现的，但你认为它是必然的吗？你认为我们可能永远无法制造出智能机器吗？

MARTIN FORD: You clearly think AGI is achievable, but do you think it’s inevitable? Do you think there is any probability that maybe we can never build an intelligent machine?

加里·马库斯：这几乎是不可避免的。我认为，阻止我们到达那里的主要因素是其他灭绝级别的生存风险，例如被小行星撞击、自我毁灭或制造超级疾病。我们正在不断积累科学知识，我们在构建软件和硬件方面也越来越好，没有原则上的理由不这样做。我认为除非我们重置时钟，否则这几乎肯定会发生，我不能排除这种可能性。

GARY MARCUS: It’s almost inevitable. I think the primary things that would keep us from getting there are other extinction-level existential risks, such as getting hit by an asteroid, blowing ourselves up, or engineering a super-disease. We’re continuously accumulating scientific knowledge, we’re getting better at building software and hardware, and there’s no principled reason why not to do it. I think it will almost certainly happen unless we reset the clock, which I can’t rule out.

马丁·福特：您如何看待国际上，特别是与中国这样的国家在先进人工智能领域的军备竞赛？

MARTIN FORD: What do you think about the international arms race toward advanced AI, particularly with countries like China?

加里·马库斯：中国已将人工智能作为其雄心壮志的主要中心，并对此非常公开。美国有一段时间没有任何回应，我发现这令人不安和沮丧。

GARY MARCUS: China has made AI a major center of its ambitions and been very public about it. The United States for a while had no response whatsoever, and I found that disturbing and upsetting.

马丁·福特：看起来中国确实有很多优势，比如人口更多、隐私控制更少，这意味着拥有更多的数据。

MARTIN FORD: It does seem that China has many advantages, such as a much larger population and fewer privacy controls, which means more data.

加里·马库斯：他们的思维更加有远见，因为他们意识到了人工智能的重要性，而且他们整个国家都在对人工智能进行投资。

GARY MARCUS: They’re much more forward-thinking because they realize how important AI is, and they are investing in it as a nation.

马丁·福特：您对该领域的监管有什么看法？您认为政府应该参与监管人工智能研究吗？

MARTIN FORD: How do you feel about regulation of the field? Do you think that the government should get involved in regulating AI research?

加里·马库斯：我明白，但我不清楚这些规定应该是什么。我认为人工智能资金的很大一部分应该解决这些问题。这些都是难题。

GARY MARCUS: I do, but it’s not clear to me what those regulations should be. I think a significant portion of AI funding should address those questions. They’re hard questions.

例如，我不喜欢自主武器的想法，但简单地彻底禁止它们可能很幼稚，而且会带来更多问题，有些人拥有它们，而其他人没有。这些规定应该是什么，我们应该如何执行它们？恐怕我没有答案。

For example, I don’t love the idea of autonomous weapons, but to simply ban them outright is maybe naive and creates more problems, where some people have them, and others don’t. What should those regulations be, and how should we enforce them? I’m afraid I don’t have the answer.

马丁·福特：您相信人工智能会给人类带来积极影响吗？

MARTIN FORD: Do you believe that AI is going to be positive for humanity?

加里·马库斯：希望如此，但我认为这不一定。人工智能帮助人类的最佳方式是加速医疗保健领域的科学发现。相反，目前的人工智能研究和实施主要是为了广告投放。

GARY MARCUS: Hopefully, but I don’t think that is a given. The best way in which AI could help humanity is by accelerating scientific discovery in healthcare. Instead, AI research and implementation right now is mostly about ad placement.

人工智能有很多积极的潜力，但我认为人们对它的关注还不够。我们做了一些，但还不够。我也明白人工智能会带来风险、失业和社会动荡。从技术角度来说，我是一个乐观主义者，因为我确实认为 AGI 是可以实现的，但我希望看到我们开发的内容和优先考虑这些事情的方式有所改变。目前，就我们如何使用人工智能以及如何分配人工智能而言，我并不完全乐观地认为我们正在朝着正确的方向前进。我认为，要让人工智能对人类产生积极影响，还有很多工作要做。

AI has a lot of positive potential, but I don’t think there’s enough focus on that side of it. We do some, but not enough. I also understand that there are going to be risks, job losses, and social upheaval. I’m an optimist in a technical sense in that I do think AGI is achievable, but I would like to see a change in what we develop and how we prioritize those things. Right now, I’m not totally optimistic that we’re heading in the right direction in terms of how we’re using AI and how we’re distributing it. I think there’s serious work to be done there to make AI have the positive impact on humanity that it could.

加里·马库斯 是纽约大学心理学和神经科学教授。加里的大部分研究都集中在了解儿童如何学习和吸收语言，以及这些发现如何为人工智能领域提供信息。

GARY MARCUS is a professor of psychology and neural science at New York University. Much of Gary’s research has focused on understanding how children learn and assimilate language, and how these findings might inform the field of artificial intelligence.

他是多部书籍的作者，包括《心灵的诞生》、《克鲁格：人类心灵的随机构造》和畅销书《吉他零》，他在书中探讨了学习弹吉他时遇到的认知挑战。加里还为《纽约客》和《纽约时报》撰写了大量关于人工智能和脑科学的文章。2014年，他创立了机器学习初创公司 Geometric Intelligence 并担任首席执行官，该公司后来被 Uber 收购。

He is the author of several books, including The Birth of the Mind, Kluge: The Haphazard Construction of the Human Mind, and the bestselling Guitar Zero, in which he explores cognitive challenges involved as he learns to play the guitar. Gary has also contributed numerous articles on AI and brain science to The New Yorker and the New York Times. In 2014 he founded and served as CEO of Geometric Intelligence, a machine learning startup that was later acquired by Uber.

Gary 以批评深度学习而闻名，他曾撰文指出，目前的方法可能很快就会“碰壁”。他指出，人类的思维并不是一张白纸，而是预先配置了重要的结构以实现学习。他认为，仅靠神经网络无法实现更通用的智能，而要继续进步，就需要将更多的先天认知结构纳入人工智能系统。

Gary is known for his criticism of deep learning and has written that current approaches may soon “hit a wall.” He points out that the human mind is not a blank slate, but comes preconfigured with significant structure to enable learning. He believes that neural networks alone will not succeed in achieving more general intelligence, and that continued progress will require incorporating more innate cognitive structure into AI systems.

哈佛大学希金斯自然科学教授

HIGGINS PROFESSOR OF NATURAL SCIENCES, HARVARD UNIVERSITY

Barbara J. Gros Barbara J. Grosz 是哈佛大学希金斯自然科学教授。在她的职业生涯中，她在人工智能领域做出了开创性的贡献，这些贡献为对话处理奠定了基础，而这些原则对于 Apple 的 Siri 或 Amazon 的 Alexa 等个人助理至关重要。1993 年，她成为第一位担任人工智能促进协会主席的女性。

Barbara J. Gros Barbara J. Grosz is Higgins Professor of Natural Sciences at Harvard University. Over the course of her career, she has made ground-breaking contributions in artificial intelligence that have led to the foundational principles of dialogue processing that are important for personal assistants like Apple’s Siri or Amazon’s Alexa. In 1993, she became the first woman to serve as president of the Association for the Advancement of Artificial Intelligence.

马丁·福特：最初是什么促使您对人工智能产生兴趣的？您的职业生涯是如何发展的？

MARTIN FORD: What initially drove you to be interested in artificial intelligence, and how did your career progress?

BARBARA GROSZ：我的职业生涯充满了各种意外的惊喜。我上大学时以为自己会成为一名七年级的数学老师，因为我的七年级数学老师是我 18 岁生命中遇到的唯一一个认为女性一般可以做数学的人，而且他还告诉我我数学很好。然而，当我上康奈尔大学时，我的世界才真正打开，因为他们刚刚开设了计算机科学系。

BARBARA GROSZ: My career was a series of happy accidents. I went to college thinking I would be a 7th-grade math teacher because my 7th-grade math teacher was the only person I had met in my first 18 years of life who thought that women, in general, could do mathematics, and he told me that I was quite good at math. My world really opened up though when I went to Cornell for college, as they had just started a computer science faculty.

当时美国没有计算机科学本科专业，但康奈尔大学提供了一些课程。我一开始学的是数值分析，这是计算机科学中一个相当数学化的领域，最后去了伯克利读研究生，最初读的是硕士学位，后来我进入了博士学位课程。

At the time there was no undergraduate major in computer science anywhere in the US, but Cornell provided the opportunity to take a few classes. I started in numerical analysis, a rather mathematical area of computer science, and ended up going to Berkeley to graduate school, initially for a master’s, then I moved into the PhD program.

我从事的是后来被称为计算科学的领域，后来又短暂从事过理论计算机科学。我认为我喜欢计算机科学数学领域的解决方案，但不喜欢问题。因此，当我需要论文主题时，我与很多人进行了交流。艾伦·凯对我说：“听着。你必须为你的论文做一些雄心勃勃的事情。你为什么不编写一个程序来阅读儿童故事并从其中一个角色的角度讲述它呢？”这激发了我对自然语言处理的兴趣，也是我成为人工智能研究员的根源。

I worked in what would come to be called computational science and then briefly in theoretical computer science. I decided that I liked the solutions in the mathematical areas of computer science, but not the problems. So when I needed a thesis topic, I talked with many people. Alan Kay said to me, “Listen. You have to do something ambitious for your thesis. Why don’t you write a program that will read a children’s story and tell it back from one of the character’s points of view?” That’s what spurred my interest in natural language processing and is the root of my becoming an AI researcher.

马丁·福特：艾伦·凯？他在施乐帕洛阿尔托研究中心发明了图形用户界面，对吧？史蒂夫·乔布斯就是在那里得到了 Macintosh 的灵感。

MARTIN FORD: Alan Kay? He invented the graphical user interface at Xerox PARC, right? That’s where Steve Jobs got the idea for the Macintosh.

BARBARA GROSZ：是的，艾伦是施乐 PARC 工作中的关键人物。我实际上与他合作开发了一种名为 Smalltalk 的编程语言，这是一种面向对象的语言。我们的目标是建立一个适合学生（K-12）和学习的系统。我的儿童故事程序将用 Smalltalk 编写。然而，在 Smalltalk 系统完成之前，我意识到儿童故事不仅仅是用来阅读和理解的故事，它们还旨在灌输一种文化，艾伦向我提出的挑战将很难应对。

BARBARA GROSZ: Yes, right, Alan was a key player in that Xerox PARC work. I actually worked with him on developing a programming language called Smalltalk, which was an object-oriented language. Our goal was to build a system suitable for students [K-12] and learning. My children’s story program was to be written in Smalltalk. Before the Smalltalk system was finished, though, I realized that children’s stories were not just stories to be read and understood, but that they’re meant to inculcate a culture, and that Alan’s challenge to me was going to be really hard to meet.

当时，第一批语音理解系统也正在通过 DARPA 项目进行开发，参与其中的 SRI International 人员对我说：“如果你愿意冒险研究儿童故事，为什么不和我们一起研究一种更客观的语言，一种以任务为导向的对话，但使用语音而不是文本？”结果，我参与了 DARPA 的语音工作，该工作是研究能够帮助人们完成任务的系统，而这也是我真正开始从事人工智能研究的时候。

During that time, the first group of speech-understanding systems were also being developed through DARPA projects, and the people at SRI International who were working on one of them said to me, “If you’re willing to take the risk of working on children’s stories, why don’t you come work with us on a more objective kind of language, task-oriented dialogues, but using speech not text?” As a result, I got involved in the DARPA speech work, which was on systems that would assist people in getting tasks done, and that’s really when I started to do AI research.

正是这项工作让我发现，当人们一起完成一项任务时，对话的结构取决于任务结构，而且对话远不止是问答对。从这一洞察中，我意识到，作为人类，我们通常不会以一连串孤立的话语说话，而是总有一个更大的结构，就像期刊文章、报纸文章、教科书，甚至这本书一样，我们可以对这个结构进行建模。这是我对自然语言处理和人工智能的第一个重大贡献。

It was that work which led to my discovery of how dialogue among people, when they’re working on a task together, has a structure that depends on the task structure—and that a dialogue is much more than just question-answer pairs. From that insight, I came to realize that as human beings we don’t in general ever speak in a sequence of isolated utterances, but that there’s always a larger structure, much like there is for a journal article, a newspaper article, a textbook, even for this book, and that we can model that structure. This was my first major contribution to natural-language processing and AI.

马丁·福特：您谈到了您最为人熟知的自然语言突破之一：努力以某种方式模拟对话。对话是可以计算的，对话中的某些结构可以用数学表示。

MARTIN FORD: You’ve touched on one of the natural language breakthroughs that you’re most known for: an effort to somehow model a conversation. The idea that a conversation can be computed, and that there’s some structure within a conversation that can be represented mathematically.

我认为这已经变得非常重要，因为我们已经看到该领域取得了很多进展。也许你可以谈谈你在那里所做的一些工作以及事情是如何进展的。与你开始研究时相比，自然语言处理方面的现状是否让你感到惊讶？

I assume that this has become very important, because we’ve seen a lot of progress in the field. Maybe you could talk about some of the work you’ve done there and how things have progressed. Has it astonished you where things are at now in terms of natural language processing, compared to where they were back when you started your research?

芭芭拉·格罗斯：这绝对让我很惊讶。我早期的工作正是在这个领域，即如何构建一个能够流畅地、自然地与人进行对话的计算机系统。我与艾伦·凯联系并和他一起工作的原因之一是，我们对构建能够与人合作并适应人的计算机系统有着共同的兴趣，而不是要求人们去适应它们。

BARBARA GROSZ: It absolutely has astonished me. My early work was exactly in this area of how we might be able to build a computer system that could carry on a dialogue with a person fluently and in a way that seemed natural. One of the reasons I got connected to Alan Kay, and did that work with him, was because we shared an interest in building computer systems that would work with and adapt to people, rather than require people to adapt to them.

在我接手这项工作的时候，语言学领域已经有很多关于句法的研究，哲学和语言学领域也有很多关于形式语义的研究，计算机科学领域也有很多关于解析算法的研究。人们知道语言理解不只是单个句子，他们也知道语境很重要，但他们没有正式的工具、数学和计算结构来将语境考虑在语音系统中。

At the time that I took that work on, there was a lot of work in linguistics on syntax and on formal semantics in philosophy and linguistics, and on parsing algorithms in computer science. People knew there was more to language understanding than an individual sentence, and they knew that context mattered, but they had no formal tools, no mathematics, and no computational constructs to take that context into account in speech systems.

我当时对人们说，我们不能只是假设正在发生的事情，我们不能只是进行内省，我们必须收集人们在执行任务时如何进行对话的样本。因此，我发明了这种方法，后来被一些心理学家称为“绿野仙踪”方法。在这项工作中，我让两个人——在这种情况下，一个专家和一个学徒——坐在两个不同的房间里，让专家向学徒解释如何完成某件事。通过研究他们一起工作产生的对话，我认识到了这些对话的结构及其对任务结构的依赖性。

I said to people at the time that we couldn’t afford to just hypothesize about what was going on, that we couldn’t just carry on introspecting, that we had to get samples of how people actually carry on a dialogue when they’re doing a task. As a result, I invented this approach, which later was dubbed the “The Wizard of Oz” approach by some psychologists. In this work, I sat two people—in this case, an expert and an apprentice—in two different rooms, and I had the expert explain to the apprentice how to get something done. It was by studying the dialogues that resulted from their working together that I recognized the structure in these dialogues and its dependence on task structure.

后来，我与 Candy Sidner 合作撰写了一篇论文，题为《注意力、意图和话语结构》。在那篇论文中，我们认为对话的结构部分是语言本身，部分是你说话的原因和说话目的的有意结构。这种有意结构是任务结构的概括。然后，这些结构方面由注意力状态模型来调节。

Later, I co-wrote a paper with Candy Sidner titled Attention, Intentions, and the Structure of Discourse. In that paper we argue that dialogues have a structure that is in part the language itself and is in part the intentional structure of why you’re speaking, and what your purposes are when speaking. This intentional structure was a generalization of task structure. These structural aspects are then moderated by a model of the attentional state.

马丁·福特：让我们快进一下，谈谈今天的情况。你看到的最大变化是什么？

MARTIN FORD: Let’s fast forward and talk about today. What’s the biggest difference that you’ve seen?

BARBARA GROSZ：我认为最大的不同是语音系统从本质上听不见的语音系统转变为如今处理语音能力极强的系统。在早期，我们无法从语音中获取太多信息，而且当时很难得到正确的解析和含义。我们也取得了长足的进步，如今的技术能够非常出色地处理单个话语或句子，这一点你可以在现代搜索引擎和机器翻译系统中看到。

BARBARA GROSZ: The biggest difference I see is going from speech systems that were essentially deaf, to today’s systems that are incredibly good at processing speech. In the early days we really could not get much out of speech, and it proved very hard to get the right kinds of parses and meaning back then. We’ve also come a long way forward with how incredibly well today’s technology can process individual utterances or sentences, which you can see in modern search engines and machine translation systems.

然而，如果你考虑任何声称可以进行对话的系统，你会发现它们本质上是行不通的。如果对话系统限制人们遵循脚本，它们似乎会做得很好，但人们并不擅长遵循脚本。有人声称这些系统可以与人进行对话，但事实上，它们真的做不到。例如，据称可以与孩子交谈的芭比娃娃是基于脚本的，如果孩子的反应方式超出设计师的预期，它就会陷入麻烦。我认为它所犯的错误实际上引发了一些严重的道德挑战。

If you consider any of the systems that purport to carry on dialogues, however, the bottom line is they essentially don’t work. They seem to do well if the dialogue system constrains the person to following a script, but people aren’t very good at following a script. There are claims that these systems can carry on a dialogue with a person, but in truth, they really can’t. For instance, the Barbie doll that supposedly can converse with a child is script-based and gets in trouble if the child responds in a way the designers didn’t anticipate. I’ve argued that the mistakes it makes actually raise some serious ethical challenges.

所有电话个人助理系统都存在类似的例子。例如，如果你问最近的急诊室在哪里，你会得到离你最近的医院的答案，但如果你问哪里可以治疗扭伤的脚踝，系统很可能只会带你到一个网页，告诉你如何治疗扭伤的脚踝。对于扭伤的脚踝来说这不是问题，但如果你问心脏病发作，因为你认为某人得过心脏病，这实际上可能会导致死亡。人们会认为，一个可以回答其中一个问题的系统也可以回答另一个问题。

Similar examples arise with all the phone personal assistant systems. For example, if you ask where the nearest emergency room is, you’ll get an answer of the nearest hospital to wherever you are when you ask, but if you ask where you can go to get a sprained ankle treated, the system is likely to just take you to a web page that tells you how to treat a sprained ankle. That’s not a problem for a sprained ankle, but if you’re asking about a heart attack because you think someone’s had one, it could actually lead to death. People would assume a system that can answer one of those questions you can answer the other.

基于数据学习的对话系统也存在相关问题。去年夏天（2017 年），我获得了计算语言学协会终身成就奖，几乎所有在会议上听我演讲的人都在研究基于深度学习的自然语言系统。我告诉他们，“如果你想建立一个对话系统，你必须认识到 Twitter 不是真正的对话。”要建立一个能够处理人们实际参与的对话的对话系统，你需要拥有真实的人进行真实对话的真实数据，而这比 Twitter 数据更难获得。

A related problem arises with dialogue systems based on learning from data. Last summer (2017), I was given the Association for Computational Linguistics Lifetime Achievement Award and almost all the people listening to my talk at the conference work on deep learning based natural-language systems. I told them, “if you want to build a dialogue system, you have to recognize that Twitter is not a real dialogue.” To build a dialogue system that can handle dialogues of the sort people actually engage in, you need to have real data of real people having real dialogues, and that’s much harder to get than Twitter data.

马丁·福特：当你谈到脱离剧本时，在我看来，这是纯语言处理和真正智能之间的模糊界限。脱离剧本和处理不可预测情况的能力才是真正的智能；这是自动机或机器人与人之间的区别。

MARTIN FORD: When you talk about going off script, it seems to me that this is the blurry line between pure language processing and real intelligence. The ability to go off script and deal with unpredictable situations is what true intelligence is all about; it’s the difference between an automaton or robot and a person.

BARBARA GROSZ：你说得完全正确，这正是问题所在。如果你拥有大量数据，那么通过深度学习，你可以将一种语言中的句子转换为另一种语言中的相同句子；或者将一个包含问题的句子转换为该问题的答案；或者将一个句子转换为可能的下一个句子，但是你无法真正理解这些句子的实际含义，因此无法根据脚本进行工作。

BARBARA GROSZ: You’re exactly right, and that’s exactly the problem. If you think about having a lot of data, that, with deep learning, enables you to, say, go from a sentence in one language to the same sentence in another language; or to go from a sentence with a question in it to an answer to that question; or from one sentence to a possible following sentence, there’s no real understanding of what those sentences actually mean, so there’s no way to work off script with them.

这个问题与保罗·格赖斯、JL·奥斯汀和约翰·塞尔在 20 世纪 60 年代提出的一个哲学观点有关，即语言就是行动。例如，如果我对计算机说：“打印机坏了”，那么我不希望它回复我：“谢谢，事实已记录。”我真正希望的是系统做一些事情来修好打印机。要做到这一点，系统需要理解我说这些话的原因。

This problem links back to a philosophical idea that was elaborated in the 1960s by Paul Grice, J. L. Austin, and John Searle that language is action. For example, if I say to the computer, “The printer is broken,” then what I don’t want is for it to say back to me, “Thanks, fact recorded.” What I actually want is for the system to do something that will get the printer fixed. For that to occur, the system needs to understand why I said something.

目前基于深度学习的自然语言系统在处理这类句子时表现不佳。原因非常深刻。我们在这里看到的是，这些系统非常擅长统计学习、模式识别和大规模数据分析，但它们不会深入研究。它们无法推断某人说话背后的目的。换句话说，它们忽略了对话的有意结构成分。基于深度学习的系统通常缺乏其他智能特征：它们无法进行反事实推理或常识推理。

Current deep-learning based natural-language systems perform poorly on these kinds of sentences in general. The reasons are really deeply rooted. What we’re seeing here, is that these systems are really good at statistical learning, pattern recognition and large-scale data analysis, but they don’t go below the surface. They can’t reason about the purposes behind what someone says. Put another way, they ignore the intentional structure component of dialogue. Deep-learning based systems more generally lack other hallmarks of intelligence: they cannot do counterfactual reasoning or common-sense reasoning.

您需要具备所有这些能力才能参与对话，除非您严格限制一个人所说和所做的事；但这会使人们很难真正做他们想做的事情！

You need all these capabilities to participate in a dialogue, unless you tightly constrain what a person says and does; but that makes it very hard for people to actually do what they want to do!

马丁·福特：您认为目前最先进的技术是什么？当我看到 IBM Watson 在《危险边缘》中获胜时，我非常惊讶！我认为这真的非常了不起。这是否像看上去的那样是一个突破，或者您认为其他技术才是真正处于领先地位的？

MARTIN FORD: What would you point to as being state-of-the-art right now? I was pretty astonished when I saw IBM Watson win at Jeopardy! I thought that was really remarkable. Was that as much of a breakthrough as it seemed to be, or would you point to something else as really being on the leading edge?

BARBARA GROSZ：Apple 的 Siri 和 IBM 的 Watson 给我留下了深刻的印象；它们是工程学的非凡成就。我认为当今的自然语言和语音系统非常棒。它改变了我们与计算机系统交互的方式，并使我们能够完成很多工作。但这些系统远不及人类的语言能力，当你尝试与它们对话时，你就会发现这一点。

BARBARA GROSZ: I was impressed by Apple’s Siri and by IBM’s Watson; they were phenomenal achievements of engineering. I think that what is available today with natural language and speech systems is terrific. It’s changing the way that we interact with computer systems, and it’s enabling us to get a lot done. But these systems are nowhere near the human capacity for language, and you see that when you try to engage in a dialogue with them.

2011 年 Siri 推出时，我大概只问了三个问题就搞定了系统。Watson 犯错的地方最有趣，因为它向我们展示了它在哪些方面不像人类那样处理语言。

When Siri came out it in 2011, it took me about three questions to break the system. Where Watson makes mistakes is most interesting in that it shows us where it is not processing language like people do.

所以是的，一方面，我认为自然语言和语音系统的进步是惊人的。我们远远超出了 70 年代所能做到的，部分原因是计算机功能强大得多，部分原因是有更多数据可用。我很高兴人工智能真的在世界上发挥作用，因为我从未想过这会在我有生之年发生——因为问题似乎太难了。

So yes, on the one hand, I think the progress in natural language and speech systems is phenomenal. We are far beyond what we could do in the ‘70s, partly because computers are way more powerful, and partly because there’s a lot more data out there. I’m thrilled that AI is actually out in the world making a difference because I didn’t think that it would happen in my lifetime—because it seemed the problems were so hard.

马丁·福特：真的吗，你没想到这会在你一生中发生？

MARTIN FORD: Really, you didn’t think it would happen in your lifetime?

芭芭拉·格罗斯：那是 1970 年代的事吗？没有。

BARBARA GROSZ: Back in the 1970s? No, I didn’t.

马丁·福特：沃森确实让我大吃一惊，尤其是它能够处理双关语、笑话以及非常复杂的语言表达。

MARTIN FORD: I was certainly very taken aback by Watson and especially by the fact that it could handle, for example, puns, jokes, and very complex presentations of language.

BARBARA GROSZ：但回到《绿野仙踪》的类比，看看这些系统背后的实际内容，你就会发现它们都有局限性。现在，我们真正需要做的是了解这些系统的优点和缺点。

BARBARA GROSZ: But just going back to “The Wizard of Oz” analogy, you look behind what’s actually in those systems, and you realize they all have limitations. We’re at a moment where it’s really important to understand what these systems are good at and where they fail.

这就是为什么我认为对于这个领域，坦率地说对于全世界来说，了解这一点非常重要，我们可以在人工智能系统方面取得更多进展，这将造福世界上的人们，如果我们的目标不是取代人类，或者建立通用人工智能，而是将我们的理解重点放在所有这些伟大的能力既有益又无益，以及如何用这些系统来补充人类，以及如何用这些系统来补充人类。

This is why I think it’s very important for the field, and frankly for the world, to understand that we could make a lot more progress on AI systems that would be good for people in the world if we didn’t aim to replace people, or build generalized artificial intelligence—but if we instead focus our understanding on what all these great capabilities are both good and not good for, and how to complement people with these systems, and these systems with the people.

马丁·福特：让我们集中讨论一下脱离剧本并真正进行对话的想法。这与图灵测试直接相关，我知道您在这方面做了一些额外的工作。您认为图灵提出这个测试的意图是什么？它是对机器智能的良好测试吗？

MARTIN FORD: Let’s focus on this idea of going off script and being able to really have a conversation. That relates directly to the Turing test, and I know you’ve done some additional work in that area. What do you think Turing’s intentions were in coming up with that test? Is it a good test of machine intelligence?

BARBARA GROSZ：我提醒大家，图灵在 1950 年提出了他的测试，当时人们拥有了他们认为非常神奇的新计算机。当然，与今天的智能手机相比，这些系统无能为力，但当时许多人想知道这些机器是否能像人类一样思考。请记住，图灵使用“智能”和“思考”的方式类似——他指的不是诺贝尔奖获奖科学类型的智能。

BARBARA GROSZ: I remind people that Turing proposed his test in 1950, a time where people had new computing machines that they thought were amazing. Now of course, those systems could do nothing compared to what a smartphone can do today, but at the time many people wondered if these machines could think like a human thinks. Remember, Turing used “intelligence” and “thinking” similarly—he wasn’t talking about intelligence like say, Nobel prize-winning science type of intelligence.

图灵提出了一个非常有趣的哲学问题，并对机器是否能够表现出某种行为做出了一些猜测。20 世纪 50 年代也是心理学扎根于行为主义的时代，因此他的测试不仅是操作性的测试，而且也是一项不会深入探究的测试。

Turing was posing a very interesting philosophical question, and he made some conjectures about whether or not machines could exhibit a certain kind of behavior. The 1950s was also at a time where psychology was rooted in behaviorism, and so his test is not only an operational test but also a test where there would be no looking below the surface.

图灵测试并不是一个好的智力测试。坦白说，我可能无法通过图灵测试，因为我不擅长社交。它也不是该领域应该努力的方向的良好指南。图灵是一个非常聪明的人，但我曾认真地推测，如果他今天还活着——如果他知道我们现在所知道的学习是如何进行的、大脑和语言是如何工作的，以及人们如何发展智力和思维，那么他会提出一个不同的测试。

The Turing test is not a good test of intelligence. Frankly, I would probably fail the Turing test because I’m not very good at social banter. It’s also not a good guide for what the field should aim to do. Turing was an amazingly smart person, but I’ve conjectured, somewhat seriously, that if he were alive today—and if he knew what we now know about how learning works, how the brain and language work, and how people develop intelligence and thinking, then he would have proposed a different test.

马丁·福特：我知道你提出了一些对图灵测试的增强，甚至是替代。

MARTIN FORD: I know that you’ve proposed some enhancements or even a replacement for the Turing test.

BARBARA GROSZ：谁知道图灵会提出什么建议，但我的建议是，鉴于我们知道人类智力的发展取决于社会互动，语言能力取决于社会互动，而人类活动在许多情况下都是协作性的——那么我建议我们致力于建立一个系统，它是一个很好的团队合作伙伴，与我们合作得很好，以至于我们不会意识到它不是人类。我的意思是，我们并不是被愚弄，以为笔记本电脑、机器人或手机是人类，而是当它犯下人类不会犯的错误时，你不会一直想“它为什么会这样做？”

BARBARA GROSZ: Who knows what Turing would have proposed, but I have made a proposal that, given that we know that the development of human intelligence depends on social interaction, and that language capacity depends on social interaction, and that human activity in many setting is collaborative—then I recommend that we aim to build a system that is a good team partner, and works so well with us that we don’t recognize that it isn’t human. I mean, it’s not that we’re fooled into the idea that a laptop, robot, or phone is a human being, but that you don’t keep wondering “Why did it do that?” when it makes a mistake that no human would.

我认为这是该领域的一个更好的目标，部分原因是它比图灵测试有几个优势。一个优势是你可以逐步实现它——所以如果你选择一个足够小的领域来构建一个系统，你可以构建一个在这个领域中智能的系统，并且它在那种任务上表现良好。我们现在可以找到我们认为以这种方式智能的系统——当然，随着儿童的成长，他们会以不同的有限方式变得聪明，然后他们会以更多样化的方式变得越来越聪明。

I think that this is a better goal for the field, in part because it has several advantages over the Turing test. One advantage is that you can meet it incrementally—so if you pick a small enough arena in which to build a system, you can build a system that’s intelligent in that arena, and it works well on that kind of task. We could find systems out there now that we would say are intelligent in that way—and of course children, as they develop, are intelligent in different limited ways, and then they get more and different kinds of smart in more varied kinds of ways.

在图灵测试中，系统要么成功，要么失败，而且没有关于如何逐步改进其推理能力的指南。科学要发展，就需要能够一步步前进。我提出的测试还认识到，在可预见的未来，人类和计算机系统将具有互补的能力，它建立在这种洞察力之上，而不是忽略它。

With the Turing test, a system either succeeds or it fails, and there’s no guide for how to incrementally improve its reasoning. For science to develop, you need to be able to make steps along the way. The test I proposed also recognizes that for the foreseeable future people and computer systems will have complementary abilities, and it builds on that insight rather than ignoring it.

我第一次提出这个测试是在爱丁堡纪念图灵诞辰 100 周年的一次演讲中。我说，鉴于计算机和心理学的所有进步，“我们应该考虑新的测试。”我在那次演讲中以及随后的演讲中询问了与会者的想法。到目前为止，主要的反应是这个测试很好。

I first proposed this test in a talk in Edinburgh on the occasion of the 100th anniversary of Turing’s birth. I said given all the progress in computing and psychology, “We should think of new tests.” I asked the attendees at that talk for their ideas, and in subsequent talks. To date, the main response has been that this test is a good one.

马丁·福特：我一直认为，一旦我们真的拥有了机器智能，我们一眼就能看出来。这在某种程度上是显而易见的，也许没有一个真正明确的测试可以定义。我不确定是否有一个测试可以测试人类智能。我的意思是，你怎么知道另一个人是聪明的？

MARTIN FORD: I’ve always thought that once we really have machine intelligence, we’ll just kind of know it when we see it. It’ll just be somehow obvious, and maybe there’s not a really explicit test that you can define. I’m not sure there’s a single test for human intelligence. I mean, how do you know another human being is intelligent?

芭芭拉·格罗斯：这真是一个非常好的观察。如果你仔细想想我举的这个例子：“最近的急诊室在哪里？我心脏病发作时可以去哪里治疗？”，你会发现没有一个你认为聪明的人能够回答这两个问题中的一个，而不能回答另一个。

BARBARA GROSZ: That’s a really good observation. If you think about what I said when I gave this example of “where’s the nearest emergency room and where can I go to get a heart attack treated?”, no human being you would consider intelligent would be able to answer one of those questions and not the other one.

有可能你问的人无法回答这两个问题，比如你把他们带到某个外国城市；但如果他们能回答一个问题，他们也能回答另一个问题。关键是，如果你有一台可以回答这两个问题的机器，那么在你看来，这台机器很聪明。如果你有一台只能回答一个问题而不回答另一个问题的机器，那么它看起来就不那么聪明了。

There’s a possibility that the person you asked might not be able to answer either question, say if you plonked them in some foreign city; but if they could answer one question, they could answer the other question. The point is, if you have a machine that answers both questions, then that seems intelligent to you. If you have a machine that answers only one and not the other question, then it doesn’t seem so intelligent.

你刚才说的其实符合我提出的测试。如果人工智能系统能够像你期望的那样聪明地运行和行动，那么你就会认为它是聪明的。现在许多人工智能系统的情况是，人们认为人工智能系统很聪明，但它却做了一些让他们大吃一惊的事情，然后他们认为它完全是愚蠢的。这时，人们想知道为什么人工智能系统会这样工作，或者为什么没有按照他们预期的方式工作，最后他们不再认为它那么聪明了。

What you just said actually fits with the test that I proposed. If the AI system is going along and acting, as it were, as intelligently as you would expect another human to act, then you’d think it is intelligent. What happens right now with many AI systems, is that people think the AI system is smart and then it does something that takes them aback, and then they think it’s completely stupid. At that point, the human wants to know why the AI system worked that way or didn’t work the way they expected, and by the end they no longer think it’s so smart.

顺便说一句，我提议的测试没有时间限制；事实上，它应该在时间上有所延长。图灵测试也不应该有时间限制，但这一特点经常被人遗忘，尤其是在最近的各种人工智能竞赛中。

By the way, the test that I proposed is not time-limited; in fact, it is actually supposed to be extended in time. Turing’s test was also not supposed to have a time limit, but that characteristic has been frequently forgotten, in particular in various recent AI competitions.

马丁·福特：这听起来很荒谬。人们的智慧不会只持续半个小时。智慧必须持续一段无限的时间才能展现出来。我认为有一种叫做洛布纳奖的东西，每年都会在某些有限条件下进行图灵测试。

MARTIN FORD: That seems silly. People aren’t intelligent for only half an hour. It has to be for an indefinite time period to demonstrate true intelligence. I think there’s something called the Loebner Prize where Turing tests are run under certain limited conditions each year.

BARBARA GROSZ：是的，这证明了你所说的。这也明确了我们在自然语言处理领域很早就学到的东西，那就是如果你只有一个固定的任务和一组固定的问题（在这种情况下，是固定的时间），那么廉价的黑客总是会胜过真正的智能处理，因为你只会根据测试来设计你的人工智能系统！

BARBARA GROSZ: Right, and it proves what you say. It also makes clear what we learned very early on in the natural-language processing arena, which is that if you have only a fixed task with a fixed set of issues (and in this case, a fixed amount of time), then cheap hacks will always win over real intelligent processing, because you’ll just design your AI system to the test!

马丁·福特：您研究的另一个领域是多智能体系统，这听起来相当深奥。您能谈谈这个领域并解释一下它的含义吗？

MARTIN FORD: The other area that you have worked in is multi-agent systems, which sounds pretty esoteric. Could you talk a little about that and explain what that means?

芭芭拉·格罗斯：当坎迪·西德纳和我开发我前面提到的意向话语模型时，我们首先尝试以同事的工作为基础，他们使用为单个机器人开发的人工智能规划模型来形式化言语行为理论的哲学研究。当我们尝试在对话的背景下使用这些技术时，我们发现它们是不够的。这一发现使我们意识到，团队合作或协作活动，或共同工作，不能简单地被描述为个人计划的总和。

BARBARA GROSZ: When Candy Sidner and I were developing the intentional model of discourse that I mentioned earlier, we first tried to build on the work of colleagues who were using AI models of planning developed for individual robots to formalize work in philosophy on speech act theory. When we tried to use those techniques in the context of dialogue, we found that they were inadequate. This discovery led us to the realization that teamwork or collaborative activity, or working together, cannot be characterized as simply the sum of individual plans.